Data processing system

Abstract

The present invention relates to a system that collects, integrates health data, action data, and communication data in multiple angles and comprehensively, performs cognitive function evaluation, risk prediction, individual optimized intervention plan generation, AI tutoring, and feedback generation using generative AI in order to eliminate anxiety of the elderly about the risk of cognitive function decline or dementia and to achieve the extension of healthy life, continuously proposes brain training optimized for each user, life habit improvement suggestions, introduction of social participation activities, and the like, and early detects slight changes in cognitive function and promotes visits to specialized institutions, thereby comprehensively supporting dementia prevention and improvement of quality of life.

Description

Technical Field

[0001] The technology disclosed herein relates to a system. Background Technology

[0002] Patent Document 1 discloses a personalized chatbot control method executed by at least one processor, the method comprising: receiving user speech; adding the user speech to a prompt containing instructions related to a chatbot role; encoding the prompt; and inputting the encoded prompt into a language model to generate chatbot speech in response to the user speech.

[0003] Patent document 1: Japanese Patent Application Publication No. 2022-180282.

[0004] The problem this invention addresses is that, among the elderly, especially couples around 80 years old, anxiety about age-related cognitive decline or the risk of developing dementia is increasing, yet opportunities for appropriate and sustained cognitive maintenance activities in daily life are scarce, and early detection of dementia is difficult. Given this situation, this invention comprehensively collects and analyzes various data on each user's health status, lifestyle habits, behavior patterns, and communication patterns. Based on high-precision prediction of individual cognitive status or risk, it automatically generates dementia prevention programs optimized for the user's interests or lifestyle. Furthermore, it monitors motivation maintenance and implementation through continuous AI-based guidance and feedback, detects subtle changes in cognitive function early, and prompts patients to seek medical attention at specialized institutions when necessary. This aims to reduce dementia risk, extend healthy lifespan, and alleviate the burden of medical expenses, contributing to a society where the elderly can maintain their individuality and live with peace of mind. Summary of the Invention

[0005] The means to solve the problem involved in this invention is to construct a system comprising a health data acquisition unit, an action data acquisition unit, a communication data acquisition unit, a data integration unit, a cognitive function assessment unit, a risk prediction unit, an intervention plan generation unit, and an AI guidance unit. The health data acquisition unit automatically acquires health-related data such as the user's vital signs, past medical history, and family medical history from external services or wearable terminals. The action data acquisition unit acquires action data such as the user's daily activity patterns, frequency of forgetfulness, implementation of social activities, amount of exercise, and frequency of outings. The communication data acquisition unit acquires communication data such as the user's conversation frequency, use of SNS or messaging applications, and the content or frequency of interpersonal communication. The data integration unit integrates this health data, action data, and communication data. The system generates an integrated dataset for each user. Based on this dataset, the Cognitive Function Assessment Department uses generative AI to analyze the user's daily conversations or behavior patterns and calculate their cognitive function score. Based on the cognitive function score and the integrated dataset, the Risk Prediction Department uses generative AI to individually predict the risk of developing dementia. Based on the risk prediction results and the user's interests or lifestyle, the Intervention Plan Generation Department uses generative AI to automatically generate an individually optimized prevention program. The AI ​​Counseling Department suggests the prevention program to the user via voice or text, monitors the user's implementation, conducts regular cognitive function assessments and provides feedback, detects subtle changes in cognitive function early, and recommends consultation with a specialist when necessary. This allows for the continuous and effective implementation of dementia prevention activities optimized for each user. Attached Figure Description

[0006] Figure 1 This is a conceptual diagram illustrating an example of the configuration of a data processing system according to the first embodiment.

[0007] Figure 2 This is a conceptual diagram illustrating an example of the main functions of the data processing apparatus and smart device according to the first embodiment.

[0008] Figure 3 This is a conceptual diagram illustrating an example of the data processing system configuration in the second embodiment.

[0009] Figure 4 This is a conceptual diagram illustrating an example of the main functions of the data processing device and smart glasses according to the second embodiment.

[0010] Figure 5 This is a conceptual diagram illustrating an example of the data processing system configuration in the third embodiment.

[0011] Figure 6 This is a conceptual diagram illustrating an example of the main functions of the data processing device and head-mounted terminal according to the third embodiment.

[0012] Figure 7This is a conceptual diagram illustrating an example of the data processing system configuration in the fourth embodiment.

[0013] Figure 8 This is a conceptual diagram illustrating an example of the functions of the main parts of the data processing device and robot according to the fourth embodiment.

[0014] Figure 9 It represents an emotion graph that maps multiple emotions.

[0015] Figure 10 It represents an emotion graph that maps multiple emotions.

[0016] Explanation of reference numerals in the attached figures Data processing systems 10, 210, 310, and 410 12 Data processing device 14 Smart devices 214 Smart Glasses 314 Head-mounted terminal 414 Robot. Detailed Implementation

[0017] Hereinafter, an example of an implementation of the system involved in this disclosure will be described with reference to the accompanying drawings.

[0018] First, let's explain the terms used in the following description.

[0019] In the following embodiments, the processor (hereinafter referred to as "processor") can be a single computing device or a combination of multiple computing devices. Furthermore, a processor can be a single computing device or a combination of multiple computing devices. Examples of computing devices include CPU (Central Processing Unit), GPU (Graphics Processing Unit), GPGPU (General-Purpose computing on Graphics Processing Units), APU (Accelerated Processing Unit), etc.

[0020] In the following implementation, the labeled RAM (Random Access Memory) is a memory that temporarily stores information and is used by the processor as working memory.

[0021] In the following embodiments, the labeled memory is one or more non-volatile storage devices used to store various programs and various parameters, etc. Examples of non-volatile storage devices include flash memory (SSD (Solid State Drive)), disk (e.g., hard disk) or magnetic tape, etc.

[0022] In the following implementation, the labeled Communication I / F (Interface) is an interface that includes a communication processor and an antenna, etc. The Communication I / F is responsible for communication between multiple computers. Examples of communication standards applicable to the Communication I / F include wireless communication standards such as 5G (5th Generation Mobile Communication System), Wi-Fi (registered trademark), or Bluetooth (registered trademark).

[0023] In the following implementation, "A and / or B" is synonymous with "at least one of A and B". That is, "A and / or B" means that it can be only A, only B, or a combination of A and B. Furthermore, in this specification, when "and / or" connects more than three items, the same approach as "A and / or B" applies.

[0024] [First Implementation] Figure 1 An example of the configuration of the data processing system 10 according to the first embodiment is shown.

[0025] like Figure 1 As shown, the data processing system 10 includes a data processing device 12 and an intelligent device 14. An example of the data processing device 12 is a server.

[0026] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0027] The smart device 14 includes a computer 36, a receiver 38, an output device 40, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the receiver 38, output device 40, camera 42, and communication I / F 44 are also connected to the bus 52.

[0028] The receiving device 38 includes a touchscreen 38A and a microphone 38B, etc., for receiving user input. The touchscreen 38A receives user input generated by contact with an indicator (e.g., a pen or finger). The microphone 38B receives user input generated by sound by detecting the user's voice. The control unit 46A in the processor 46 sends data representing user input received through the touchscreen 38A and microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the data representing the user input.

[0029] The output device 40 includes a display 40A and a speaker 40B, and presents data to the user 20 by outputting data in a form perceptible to the user 20 (e.g., sound and / or text). The display 40A displays visual information such as text and images according to instructions from the processor 46. The speaker 40B outputs sound according to instructions from the processor 46. The camera 42 is a small digital camera equipped with an optical system such as a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor.

[0030] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for sending and receiving various information between processor 46 and processor 28 via network 54.

[0031] Figure 2 An example of the main functions of the data processing device 12 and the smart device 14 is shown.

[0032] like Figure 2 As shown, in the data processing apparatus 12, specific processing is performed by the processor 28. A specific processing program 56 is stored in the memory 32. The specific processing program 56 is an example of a "program" as understood in this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 based on the specific processing program 56 executed on the RAM 30.

[0033] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. The emotion inference function using the emotion-specific model 59 (emotion-specific function) includes various inferences and predictions related to the user's emotions, but is not limited to this example. Furthermore, emotion inference and prediction also include, for example, emotion analysis (analysis).

[0034] In the smart device 14, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The receive / output program 60 is used in conjunction with the data processing system 10 and the specific processing program 56. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 acting as a control unit 46A based on the receive / output program 60 executed on the RAM 48. Furthermore, the smart device 14 may also have the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and use these models to perform the same processing as the specific processing unit 290. The receive / output processing is implemented by the processor 46 acting as a control unit 46A based on the receive / output program 60 executed on the RAM 48.

[0035] In addition, other devices besides the data processing device 12 may also have a data generation model 58.

[0036] For example, a server device (e.g., a generation server) may have a data generation model 58.

[0037] In this case, the data processing device 12 obtains the processing results (prediction results, etc.) using the data generation model 58 by communicating with a server device that has the data generation model 58. Furthermore, the data processing device 12 can be a server device or a user-owned terminal device (e.g., a mobile phone, robot, home appliance, etc.). Next, an example of the processing performed by the data processing system 10 of the first embodiment will be described.

[0038] (Example 1) The specific processing flow in Example 1 will be described below. Each part of the system described below is implemented by the data processing device 12 and the smart device 14. Furthermore, the data processing device 12 is referred to as the "server" and the smart device 14 is referred to as the "terminal".

[0039] (Implementation methods for carrying out the present invention) The embodiments used to carry out the present invention will be described in a more detailed and comprehensive manner to further enhance the specificity.

[0040] The system of this invention is constructed by connecting a server and user terminals (smartphones, tablets, personal computers, wearable devices, etc.) via a network. Among the various parts of the system, the health data acquisition unit, action data acquisition unit, communication data acquisition unit, data integration unit, cognitive function assessment unit, risk prediction unit, intervention plan generation unit, and AI guidance unit are mainly implemented on the server, but some data acquisition, user interface, voice input / output, notification, and feedback transmission are also performed on the user terminal side. Specifically, this system adopts a distributed processing architecture connecting the server and user terminals via a network. The system configures a parallel computing cluster using GPUs on the server side. The system installs an edge computing module on the user terminal side. The system processes the raw data acquired on the user terminal side only once at the edge. The system only sends the feature vector extracted through this single processing to the server. This structure significantly reduces the load on communication bandwidth. The system optimizes the consumption of computing resources on the server side. The system achieves real-time data processing. This technology improves and increases the overall processing speed of the system. The system provides a latency-free user experience. As a specific application area, this system is used in elderly care services or remote medical monitoring systems. This system containerizes the various processing units running on the server as a microservice architecture. Communication between these processing units is performed asynchronously via a lightweight messaging protocol. This ensures the system's scalability. A portion of the data acquisition or user interface processing is performed by the user terminal's local processor. This provides fault tolerance by offering minimal functionality even when the network is down.

[0041] First, the health data acquisition department automatically acquires the user's vital signs information (e.g., heart rate, blood pressure, body temperature, steps, sleep duration, weight, blood glucose level, oxygen saturation, calories burned, resting heart rate, exercise intensity, wake-up and bedtime times, etc.), past medical history (e.g., stroke, heart disease, diabetes, hypertension, depression, history of fractures, surgical history, etc.), family medical history (e.g., whether there are relatives with dementia, stroke, Parkinson's disease, etc.), medication history (e.g., antihypertensive drugs, antidepressants, sleeping pills, supplements, etc.), allergy information, and regular physical examination results (e.g., blood test values, urine test values, image diagnostic results, etc.) from health management applications or wearable devices (e.g., smartwatches, activity meters, blood pressure monitors, thermometers, pulse oximeters, blood glucose meters, etc.) installed on the user's terminal. This data is sent from the terminal to the server via encrypted communication and is stored and managed in a time-series database on the server side. For example, data such as daily morning blood pressure measurements, weekly step counts, and monthly weight changes are recorded at a fine granular level. Specifically, this health data acquisition department constructs a data pipeline for continuously collecting time-series data from IoT devices. This health data acquisition department digitizes the biological signals acquired from each device at a specified sampling rate. This health data acquisition department expands the digitized data as a time-series tensor in memory. This health data acquisition department performs preprocessing on this time-series tensor. As preprocessing, this health data acquisition department performs high-frequency noise removal using a moving average filter. As preprocessing, this health data acquisition department performs interpolation for missing values ​​using a spline interpolation algorithm. This health data acquisition department stores the preprocessed data in a time-series database on the server side. In the time-series database, this health data acquisition department uses a columnar format with timestamps as the primary key. This health data acquisition department thus significantly improves data writing and retrieval speeds. This health data acquisition department improves and optimizes the management efficiency of massive vital sign data through this technology. This health data acquisition department provides the foundation for subsequent AI models to extract high-precision features. This health data acquisition department, as a specific application area, is utilized in chronic disease early warning systems or personalized health management platforms. This department stores static data such as regular physical examination results as structured data in a relational database. This department also manages dynamic time-series data and static structured data in a unified manner by linking them through user IDs.

[0042] The Action Data Acquisition Department uses the user terminal's GPS function or accelerometer, gyroscope, calendar application, action recording application, voice input application, image recognition application, etc., to acquire the user's daily action patterns (e.g., frequency of going out, distance traveled, amount of exercise, number of times participating in shopping or hobby activities, completion of housework, sleep rhythm, meal times, bathing times, number of toilet visits, TV viewing time, reading time, whether or not there is a nap, diversity of walking routes, use of transportation, etc.), frequency of forgetfulness (e.g., forgetting scheduled plans, forgetting personal belongings, forgetting appointments, incorrect cooking steps, forgetting shopping lists, forgetting to take out the trash, forgetting to take medication, etc.), and completion of social activities (e.g., participating in community activities, volunteer activities, attending interest groups, visiting friends' homes, community association activities, travel, dining out with family, attending weddings and funerals, etc.). For example, it can link with calendar applications to automatically record the input, execution, and forgetting of scheduled plans, or automatically determine the destination or movement route from GPS data, or analyze the intensity of exercise or walking patterns from the accelerometer. Furthermore, using the image recognition application, it is also possible to manage food items inside the refrigerator, identify medication packaging, and confirm garbage disposal status. Specifically, this action data acquisition unit acquires multidimensional physical quantity data from various sensors mounted on the user terminal. This action data acquisition unit converts the acquired physical quantity data into a multidimensional vector. This action data acquisition unit utilizes a convolutional neural network in the image recognition application. This action data acquisition unit uses RGB image tensors acquired from a camera as input data to this neural network. This action data acquisition unit defines the dimension of the input data as a 3D array of width, height, and color channels. This action data acquisition unit performs image normalization and resizing as preprocessing of the input data. This action data acquisition unit takes as a specific example the input data, such as image tensors of food items inside the refrigerator and image tensors of medication packaging as input data. This action data acquisition unit acquires output data from the neural network. This action data acquisition unit outputs a predefined probability distribution for each object category as output data. This action data acquisition unit is configured such that each output value represents the probability of the presence of a specific object. As a specific example of output data, the action data acquisition unit outputs a probability of 0.95 for the presence of milk and 0.88 for the presence of antihypertensive drug packaging. This output result is used for subsequent processing. As part of the subsequent processing, the action data acquisition unit records the tags of objects with probabilities exceeding the threshold in the database. The action data acquisition unit utilizes a recurrent neural network in time-series action recognition. It inputs time-series data from an accelerometer into the recurrent neural network to classify action states such as walking or stationary states. Through this non-inertial step, the action data acquisition unit achieves high-precision action tracking without relying on human visual observation.The data acquisition department of this operation is used as a specific application area in the life rhythm monitoring system or rehabilitation support system for elderly people living alone.

[0043] The Communication Data Acquisition Department uses user terminal call history, messaging application usage, SNS usage, voice recognition functions, video calling applications, email applications, camera applications, etc., to obtain user conversation frequency (e.g., number of phone calls with family or friends, number of messages sent and received, frequency of SNS posts, number of video calls, number of emails sent, number of photos or videos shared, etc.) and the content or frequency of interpersonal communication (e.g., conversation length, topic diversity, presence of emotional expression, repetition of the same topic, response speed to questions, frequency of topic switching, presence of jokes or laughter, mispronouncing the other person's name, confusion in family structure, confusion of dates or days of the week, etc.). For example, the voice recognition function is used to transcribe daily conversations, and natural language processing technology is used to analyze the fluency or vocabulary diversity, logical connections, emotional expression, topic changes, consistency of conversation, and expressions of care for the other party. From the usage of SNS or messaging applications, the frequency or time period of communication, response speed, tendency to not reply to read messages, and frequency of use of stickers or emoticons can also be obtained. Specifically, this communication data acquisition unit installs a natural language processing pipeline for parsing speech and text data. In the speech recognition function, this communication data acquisition unit extracts Mel-frequency cepstral coefficients from the acquired speech waveform. This communication data acquisition unit inputs the extracted acoustic features into a Transformer-based speech recognition model, converting them into text data. This communication data acquisition unit inputs the converted text data into a large language model. This communication data acquisition unit uses word embedding vectors as input data to the large language model. This communication data acquisition unit defines the dimension of the input data as, for example, a dense vector of 768 dimensions. This communication data acquisition unit includes the semantic expression of the spoken content as the element content of the input data. This communication data acquisition unit performs stop word removal or tokenization as preprocessing of the input data. Specific examples of input data include the embedding vectors of dialogue text with a family and the embedding vectors of posts submitted to SNS. This communication data acquisition unit acquires output data from the large language model. This communication data acquisition unit outputs numerical vectors representing sentiment scores or fluency indicators as output data. This communication data acquisition unit is configured such that each output value signifies the user's psychological state or language ability level. As a specific example of output data, this communication data acquisition unit outputs a positive affect score of 0.8 and a lexical diversity score of 0.6. This communication data acquisition unit utilizes these output results for subsequent processing. As a subsequent processing step, this communication data acquisition unit performs the process of accumulating these scores over time and sending them to the data integration unit. This communication data acquisition unit detects subtle cognitive changes that are difficult for humans to perceive through operations in high-dimensional space. As a specific application area, this communication data acquisition unit is used in mental health care applications or early dementia detection systems.

[0044] The data integration department aggregates health data, activity data, and communication data for each user on the server, generating an integrated dataset managed by time series. For example, it summarizes a day's vital signs, travel history, conversation content, social activity records, medication status, meal content, sleep duration, exercise volume, SNS usage, communication with family, and the implementation of hobbies into a single record, enabling the tracking of long-term changes or trends. The data integration department also performs data correction for missing or outlier values, anonymization for privacy protection, and data normalization or standardization. Specifically, this data integration department executes early or late fusion algorithms for integrating heterogeneous multimodal data. This data integration department combines feature vectors extracted from each of the health data, activity data, and communication data in memory. This data integration department generates a high-dimensional integrated feature vector. This data integration department executes a normalization algorithm during the data integration process. As a normalization algorithm, this data integration department applies Z-score normalization or min-max scaling. This data integration department thereby unifies the scale of each feature. This data integration department uses a machine learning-based multiple imputation method to accurately estimate and fill in missing values ​​for missing data. For privacy protection, it uses an irreversible hash function to anonymize personally identifiable information. The anonymized data is stored in a secure storage area. These processes significantly improve the learning and convergence speed of subsequent machine learning models. This data integration department delivers improved computational efficiency and model prediction accuracy. As a specific application area, this data integration department is used in large-scale medical data analysis platforms or personal health record management systems. It can also store the integrated dataset in a graph database, using nodes and edges to represent the relationships between data points. This enables high-speed retrieval and analysis of complex coherent relationships.

[0045] The Cognitive Function Assessment Department parses and integrates datasets on a server, utilizing generative AI to calculate cognitive function scores from users' daily conversations and behavior patterns. For example, it uses generative AI-based natural language processing technology to analyze conversation content in detail, incorporating indicators such as repetition of the same topic or pauses in speech, response speed to questions, vocabulary diversity, logical coherence, richness of emotional expression, use of jokes or metaphors, accurate recognition of the other person's name or date, consistency of conversation, frequency of topic switching, topic deviation, and expressions of care for the other person. Furthermore, changes in behavior patterns (e.g., a sharp decrease in the frequency of going out, reduced willingness to participate in activities of interest, increased forgetfulness, disrupted sleep rhythms, reduced mealtimes, increased forgetfulness of taking medication, and errors in household chores) are also reflected in the score. Further, using generative AI, it extracts subtle changes in cognitive function from the user's overall life and quantifies them into scores out of 100. For example, weekly score shifts are graphically plotted and compared with past averages or standard deviations to detect outliers. Specifically, this cognitive function assessment unit utilizes a large language model based on the Transformer architecture to calculate cognitive function scores. This cognitive function assessment unit inputs data into the large language model. As input data, this cognitive function assessment unit uses cross-modal feature vectors generated in the data integration unit. This cognitive function assessment unit sets the dimension of the input data to, for example, 1024 dimensions. The elemental content of the input data includes embedded expressions of the dialogue text, numerical arrays of action frequencies, and time-series statistics of vital sign data. As preprocessing of the input data, this cognitive function assessment unit performs a linear transformation process that projects the data from each modality to the same latent space. As a specific example of input data, this cognitive function assessment unit inputs a numerical vector representing the number of outings in the past week and a time-series array representing the speech intervals in the dialogue. This cognitive function assessment unit obtains output data from the large language model. As output data, this cognitive function assessment unit outputs a continuous score from 0 to 100 and a probability distribution for each cognitive domain. This cognitive function assessment unit is configured such that each output value signifies the user's current overall cognitive function level, or the risk of decline in specific functions such as memory or language ability. As a specific example of output data, this cognitive function assessment unit outputs a comprehensive cognitive score of 85 and a floating-point number indicating a language ability decline probability of 0.2. This output result is used in subsequent processing. As part of this subsequent processing, the unit inputs the output score into a time-series anomaly detection algorithm to calculate the deviation from past baselines. This unit uses self-supervised learning and a cross-entropy error function to pre-learn the model. Through this technique, the unit achieves an objective and highly accurate assessment that is independent of human subjectivity.As a specific application area, this cognitive function assessment department is used in online cognitive function testing services or smart home systems for the elderly.

[0046] The Risk Prediction Department uses a server-side system based on cognitive function scores and integrated datasets to individually predict the risk of developing dementia using generative AI-based machine learning models (e.g., deep learning models, decision tree models, random forest models, gradient boosting models, etc.). For example, it takes multiple factors as input, including family history or past medical history, vital signs, changes in cognitive function scores, frequency of social activities, quality of communication, lifestyle habits (e.g., smoking, drinking, exercise, eating, sleep, stress, hobbies, social participation, intellectual activities, emotional expression, etc.), medication status, medical records, fall history, accident history, driving history, economic status, living environment, and support systems. Based on patterns learned by generative AI from vast amounts of past data, it calculates the risk level (e.g., low, medium, high). Furthermore, the generative AI can also explain the risk factors to users or their families in natural language, or which factor contributes to the increased risk. For example, it could specifically state, "Due to recent reduced outings and increased repetitive conversations, the risk of dementia is increasing." Specifically, this risk prediction unit employs an ensemble learning architecture that combines a gradient boosting model or a deep learning model. This risk prediction unit inputs data into the machine learning model. As input data, this risk prediction unit uses time-series trending data of integrated feature vectors and cognitive function scores. This risk prediction unit sets the dimension of the input data to, for example, 2048 dimensions. The element content of the input data includes numerical vectors of age or blood pressure, arrays of social activity frequencies, and genetic risk scores. As preprocessing of the input data, this risk prediction unit performs dimensionality reduction based on principal component analysis. As a specific example of input data, this risk prediction unit inputs an array representing the trend of cognitive scores over the past 6 months and a numerical value representing the coefficient of variation of sleep duration. This risk prediction unit obtains output data from the machine learning model. As output data, this risk prediction unit outputs a score representing the probability of developing dementia from 0.0 to 1.0, and a SHAP value representing the risk contribution of each feature. This risk prediction unit is structured so that each output value signifies the statistical probability of developing dementia in a specific future period, and the specific factors that increase that probability. As a specific example of output data, this risk prediction department outputs a 3-year incidence rate of 0.75 and a contribution of reduced outings of 0.3. This risk prediction department utilizes this output result for subsequent processing. As part of the subsequent processing, this risk prediction department executes a branch process that sends a trigger signal to the intervention plan generation department when the incidence rate exceeds a predetermined threshold. In model learning, this risk prediction department uses gradient descent and backpropagation algorithms to optimize weights. This risk prediction department eliminates the black-box problem of AI through the calculation of SHAP values, ensuring the transparency of the prediction basis. As a specific application area, this risk prediction department is used in preventive healthcare support systems or health improvement programs of insurance companies.

[0047] The intervention plan generation department uses generative AI on the server to automatically generate individually optimized prevention programs based on risk prediction results and users' interests or lifestyles. For example, if a user is interested in puzzles or music, cooking, gardening, reading, travel, photography, handicrafts, Go, Shogi, karaoke, walking, dancing, yoga, volunteer activities, community activities, communication with family, or pet ownership, the generative AI will combine these interests with brain training programs (e.g., crossword puzzles, Sudoku, memory games, calculation problems, word games, illustration memorization, story creation, photo classification, music appreciation, singing, instrument playing, cooking recipe memorization, gardening diary creation, travel journal creation, book review creation, photo album creation, handicraft record creation, Go and Shogi games, karaoke scoring, walking route suggestions, dance video viewing, yoga posture guidance, introduction to volunteer activities, community activity participation guides, suggestions for common topics with family, suggestions for playing with pets, etc.) and lifestyle habit improvement suggestions (e.g., The proposals cover a variety of interventions, including: balanced diet, exercise habits, sleep rhythm, stress management, medication management, smoking cessation and alcohol limitation, fall prevention, infection prevention, oral care, excretion management, bathing habits, handwashing and rinsing, weight management, blood pressure management, blood sugar management, supplement utilization, and recommendations for health check-ups and medical appointments. They also introduce social participation activities (e.g., community groups, interest classes, sports clubs, volunteer groups, community association activities, tours, cultural festivals, lectures, film screenings, karaoke parties, cooking classes, gardening classes, book clubs, photography exhibitions, Go and Shogi tournaments, pet activities, etc.) and recommendations for specialized institutions (e.g., dementia specialists, neurology specialists, psychiatrists, psychosomatic medicine specialists, rehabilitation specialists, community comprehensive support centers, nursing professionals, visiting nurses, pharmacist consultations, nutritionist consultations, social welfare worker consultations, etc.). Furthermore, generative AI-based image generation technology can be used to automatically generate user-specific brain training illustrations or quizzes, and personalized cognitive stimulation content (e.g., memory training using family photos, quizzes about birthplaces or places of memory, illustrations or videos about interests, and brain training content tailored to seasons or activities). Specifically, this intervention plan generation unit employs a hybrid architecture integrating reinforcement learning algorithms and diffusion-based image generation AI. This intervention plan generation unit inputs data into the reinforcement learning model. As input data, this intervention plan generation unit uses the user's profile vector, risk prediction score, and past intervention history tensor. This intervention plan generation unit defines the dimension of this input data as a dynamically variable tensor. As element content of the input data, this intervention plan generation unit includes categorical variables representing the user's interests and hobbies, and numerical arrays representing past task completion rates. As preprocessing of the input data, this intervention plan generation unit performs one-hot encoding of the categorical variables.As a specific example of input data, the intervention plan generation unit takes as input a vector representing a preferred music and a numerical value representing a past puzzle completion rate of 0.8. The intervention plan generation unit obtains output data from a reinforcement learning model. As output data, the intervention plan generation unit outputs an array of identifiers for the optimal intervention content and a timetable matrix. The intervention plan generation unit is structured so that each output value signifies the specific type of program to be provided to the user, and the optimal date and time for prompting that program. As a specific example of output data, the intervention plan generation unit outputs the identifier "Brain Training ID102" and a time vector representing the recommended time period. The intervention plan generation unit utilizes this output result for subsequent processing. As subsequent processing, the intervention plan generation unit performs the process of sending the generated timetable to the AI ​​tutoring unit. The intervention plan generation unit further inputs text prompt vectors into the image generation AI, causing it to output a user-specific brain training image tensor. Through this technology, the intervention plan generation unit achieves a leap in personalized accuracy and efficient allocation of computational resources. As a specific application area, the intervention plan generation unit is utilized in digital therapy applications or entertainment automation generation systems for nursing facilities.

[0048] The AI ​​Coaching Department distributes preventative programs generated on the server to user terminals. On these terminals, users receive suggestions through interactive interfaces (such as voice assistants, chatbots, notification applications, calendar integration, reminders, push notifications, vibration notifications, on-screen displays, video playback, and music playback). For example, specific action proposals might be offered daily, such as "Do 10 minutes of puzzle brain training today," "Would you like to join a local community activity this afternoon?", "You've been going out less lately, how about a walk to refresh your mood?", "Have a video call with your family today," "Would you like to try a new cooking recipe?", or "There's a local film screening this week." When users follow these proposals, their implementation is fed back to the server. The AI ​​Coaching Department uses generative AI to analyze the user's motivation and implementation, providing encouragement, reminders, or feedback at the optimal time and in the best way. For example, if a user fails to maintain their activity, the approach might be changed, such as "Don't force yourself, continue within your capabilities," "Would you like to try something with your family?", "Listen to your favorite music today to refresh your mood," "You worked very hard yesterday," or "Try something new this week." Furthermore, it also analyzes user reactions or emotional expressions, activity implementation, lifestyle, changes in physical condition, and feedback from family to dynamically optimize tutoring content. Specifically, this AI tutoring department has a dialogue generation engine using a large language model and a speech synthesis module. This AI tutoring department inputs data into the dialogue generation engine. As input data, this AI tutoring department uses the user's speech text vector and the current context vector. This AI tutoring department defines the dimension of this input data as a matrix dependent on the dimension of the labeled embedding space. As elements of the input data, this AI tutoring department includes the text expression immediately preceding the user's speech, the current time, and the user's fatigue level estimate. As preprocessing of the input data, this AI tutoring department performs contextual information cues and processing. As a specific example of input data, this AI tutoring department inputs the text vector of the user saying "tired" and the time vector of the current time "3 PM". This AI tutoring department obtains output data from the dialogue generation engine. As output data, this AI tutoring department outputs the labeled probability distribution of the response text. This AI tutoring department constructs each output value to represent the likelihood of the next word to be generated. As a specific example of output data, this AI tutoring unit outputs the text string "Don't push yourself, take a break." This output is then used for subsequent processing. As part of this subsequent processing, the AI ​​tutoring unit sends the generated text to the speech synthesis module, generating natural speech waveform data which is then output from the speaker. The AI ​​tutoring unit sets the user's proposal execution rate as the reward function and continuously updates the weights of the dialogue strategy through reinforcement learning. Through this technology, the AI ​​tutoring unit improves user engagement and achieves dynamic state adaptation.This AI tutoring department, as a specific application area, is utilized as a virtual health assistant or an interactive companion robot for the elderly.

[0049] In addition, the AI-assisted counseling department regularly re-evaluates cognitive function and uses generative AI to analyze changes in scores or behavior patterns. For example, in cases of a sharp decline in cognitive function scores, or when abnormal patterns are detected in conversation content (e.g., increased cryptic speech, delayed responses to questions, repetition of the same topics, confusion about dates or days of the week, confusion about family composition, loss of emotional expression, abandonment of hobbies, refusal to go out, reduced mealtimes, increased medication forgetfulness, increased falls, etc.), the generative AI automatically summarizes the content and sends alerts to the user or their family such as "Recommend to seek treatment at a specialized institution," "Consult your attending physician," or "Contact the local comprehensive support center." Furthermore, with the user's consent, it can also collaborate with attending physicians or care providers, the local comprehensive support center, family members, social welfare personnel, nursing managers, pharmacists, nutritionists, rehabilitation personnel, etc., to share necessary information. Specifically, this AI-assisted counseling department installs a time-series anomaly detection algorithm to perform regular re-evaluations. This AI-assisted counseling department utilizes an autoencoder architecture as the anomaly detection model. This AI-assisted counseling department inputs data into the anomaly detection model. This AI tutoring department uses the latest cognitive scores and time-series tensors of action features as input data. The dimension of this input data is defined as a two-dimensional tensor consisting of the product of the number of time steps and the number of features. The elements of the input data include daily score changes or the frequency of specific actions. As preprocessing of the input data, this AI tutoring department performs data segmentation based on a sliding window approach. As a specific example of the input data, this AI tutoring department inputs an array of cognitive scores from the past 7 days or a vector of values ​​representing the number of times the same topic was repeated each day. This AI tutoring department obtains output data from an anomaly detection model. As output data, this AI tutoring department outputs the reconstruction error score and anomaly flag of the input data. This AI tutoring department is configured such that each output value indicates how much the current user state deviates from the normal pattern and whether an emergency response is needed. As a specific example of the output data, this AI tutoring department outputs a reconstruction error score of 0.85 or an anomaly flag of 1. This AI tutoring department utilizes this output result for subsequent processing. As subsequent processing, when the anomaly flag is 1, this AI tutoring department executes the process of activating the alarm generation module and sending an API call to an external medical institution system. This AI-assisted system improves real-time anomaly detection and rapid fail-safe mechanisms through this technology. As a specific application area, this AI-assisted system is utilized in remote medical monitoring systems or emergency notification services.

[0050] In this way, the system of the present invention leverages generative AI on the server side for advanced data analysis, risk prediction, intervention plan generation, and guidance control, while on the user terminal side it handles data acquisition, user interface, voice input / output, notifications, and feedback transmission. This allows for continuous and effective dementia prevention activities optimized for each user. For example, if a user enjoys traveling, generative AI can automatically generate memory training materials related to travel or brain training using travel photos, social activity proposals at the travel destination, travel log creation, quizzes on the history or culture of the travel destination, and communication enhancement with travel companions, all naturally integrated into the user's life. For instance, it can provide personalized recipe memory quizzes or challenges to new dishes to users who enjoy cooking, guides to cooking classes, album creation of cooking photos, proposals for dishes to serve to family or friends, advice on ingredient management or nutritional balance, and recipe proposals using seasonal ingredients. For example, for users who have reduced conversations with their families, the system provides reminders to encourage video calls or message exchanges with their families, proposes content on common family topics, provides memory training using family photos, proposes to celebrate family birthdays or anniversaries, and proposes joint family activities (e.g., cooking, gardening, travel, games, reading, photography, etc.), thus supporting the activation of communication. Specifically, this system builds an end-to-end data pipeline and executes distributed processing based on a multi-agent system. This system enables each processing unit to function as an independent agent, exchanging messages asynchronously between agents. This architecture improves the overall scalability of the system. Even with an increase in the number of users, the system dynamically distributes the computational load, reducing overall system latency. This system performs generative AI inference processing at high speed on dedicated GPU instances in the cloud. This system introduces a federated learning approach to coordinate lightweight inference processing on the user terminal side and heavyweight inference processing on the cloud side. This improves the accuracy of the global model while keeping user privacy data within the terminal. This technology improves both data security and personalized accuracy. This system, as a specific application, is used in nationwide healthy life extension projects or municipal-level elderly support platforms. When dynamically generating content that matches users' interests and lifestyles, the system utilizes a hybrid recommendation system that combines content-based filtering and collaborative filtering algorithms. This allows the system to uncover users' potential interests and encourage continued use.

[0051] Furthermore, generative AI also analyzes in real time changes in users' lifestyles or hobbies, the progress of implementation, seasons or weather, the status of local events, changes in physical condition or mood, the situation of family or friends, social news or topics, and local medical and nursing resources, dynamically optimizing intervention plans or guidance content. For example, it can strengthen physical condition management or infection prevention during seasonal transitions, encourage participation during local events, propose joint activities with family members when family visits are scheduled, automatically propose activities or recommend rest within a reasonable range when users report feeling unwell, and guide them to medical institutions as needed. For example, in the event of a user's fall, it can record detailed information about the fall, and can also incorporate fall prevention programs or rehabilitation proposals, prompting emergency responses such as contacting family or medical institutions. Specifically, this generative AI utilizes a streaming data processing architecture to perform real-time analysis. This generative AI continuously receives information from environmental sensor data or external APIs via an event-driven messaging system. This generative AI inputs data to the streaming processing engine. This generative AI uses real-time environmental sensor data or text vectors from news sources as input data. The dimension of this input data is defined as a variable-length tensor corresponding to the batch size of the stream. The elements of the input data include current temperature and humidity data or text information about local events. Preprocessing of the input data involves noise filtering and text vectorization. Specific examples of input data include numerical data representing a temperature of 35 degrees Celsius or an embedding vector of the text "Summer Festival". Output data is obtained from the stream processing engine. The output data is a differential vector representing an updated intervention plan. Each output value signifies a change to the existing schedule or an additional intervention task. Specific examples of output data include a cancellation indicator for outdoor activities or an additional task vector for hydration reminders. This output is used for subsequent processing. Subsequent processing involves asynchronously updating the schedule matrix in the database. This generative AI improves upon existing technologies to ensure the system's immediate responsiveness to environmental changes and to balance computational load. As a specific application area, this generative AI is utilized in smart city elderly care infrastructure or weather data-driven health management applications.

[0052] Thus, the embodiments of the present invention possess a multifaceted and robust structure, including server-user terminal linkage, multi-layered and dynamic utilization of generative AI, optimized intervention and continuous support for each user, real-time data analysis and feedback, linkage with medical, nursing, family, and regional resources, emergency response, privacy protection, and enhanced user experience. This comprehensively achieves dementia prevention, cognitive function maintenance, extended healthy lifespan, reduced medical expenses, improved quality of life for the elderly, reduced burden on families or regional communities, and improved overall social welfare. Specifically, the system involved in this embodiment is installed in a cloud computing environment as a containerized microservice architecture. The system deploys each functional module as an independent container and uses orchestration tools to dynamically manage resources. This ensures high system availability, achieving fault tolerance so that even if some modules fail, the entire system can automatically recover without stopping. The system employs end-to-end encryption protocols in data transmission and reception, completely preventing data leakage along the communication path. The system introduces a role-based access control model in database access control, blocking access from unauthorized users or processes. This system leverages these technological improvements to meet the stringent security requirements necessary for processing medical information. It maximizes the utilization efficiency of hardware resources and reduces operating costs. As a specific application area, this system is utilized in national-level medical data infrastructures or globally deployed healthcare cloud services. The system constructs an MLOps pipeline that automates the continuous data collection and model relearning cycle, keeping the inference accuracy of the AI ​​model always up-to-date.

[0053] (System Composition) The system involved in this implementation includes a health data acquisition department, an action data acquisition department, a communication data acquisition department, a data integration department, a cognitive function assessment department, a risk prediction department, an intervention plan generation department, an AI tutoring department, a user interface department, a feedback management department, and an external linkage department.

[0054] The Health Data Acquisition Unit automatically acquires various health-related data from wearable devices such as smartwatches, activity meters, blood pressure monitors, thermometers, pulse oximeters, and blood glucose meters, as well as from health management applications, electronic medical records from regular physical examinations, and medication management applications installed on smartphones or tablets. This data includes heart rate, blood pressure, body temperature, steps, sleep duration, weight, blood glucose levels, oxygen saturation, calories burned, resting heart rate, exercise intensity, wake-up and bedtime times, past medical history, family medical history, medication history, allergy information, and results of regular physical examinations. For example, it acquires daily morning blood pressure measurements from a smartwatch via Bluetooth, weekly step counts from an activity meter, monthly weight changes from a scale via Wi-Fi, and medication status from a medication management application. Data is collected from multiple devices or applications in a time-series manner and sent to the server. Specifically, this Health Data Acquisition Unit executes data compression algorithms on the edge device side. The Health Data Acquisition Unit inputs data to the microcontroller of the edge device. This health data acquisition unit uses raw waveform data acquired from biosensors as input data. The dimension of this input data is defined as a one-dimensional array with a high sampling rate. The elements of the input data include continuous data such as voltage values ​​or accelerations from an electrocardiogram (ECG). As preprocessing of the input data, the unit performs analog-to-digital conversion and bandpass filtering. As a specific example of input data, the unit inputs an array of heart rate waveforms sampled at 100Hz or continuous blood pressure variation data. The unit acquires output data from a microcontroller. As output data, the unit outputs a compressed vector after feature extraction. The unit is configured such that each output value represents statistical features such as peak intervals or frequency components extracted from the raw waveform. As a specific example of output data, the unit outputs a numerical value representing the average heart rate (72) or a floating-point number representing the standard deviation of heart rate variation. The unit utilizes this output result for subsequent processing. As subsequent processing, the unit performs a process that sends the compressed vector to a server using a secure communication protocol. This health data acquisition unit improves upon communication bandwidth and reduces battery consumption of edge devices through technological advancements. Specifically, it is used in data collection systems for wearable medical devices or sports science applications requiring long-term continuous monitoring.

[0055] The Action Data Acquisition Unit uses the user terminal's GPS function or accelerometer, gyroscope, calendar application, action recording application, voice input application, image recognition application, etc., to acquire daily life action patterns such as frequency of outings, distance traveled, amount of exercise, number of times shopping or participating in hobbies, completion of housework, sleep rhythm, meal times, bathing times, number of toilet visits, TV viewing time, reading time, whether or not there is a nap, diversity of walking routes, and use of transportation. For example, in conjunction with the calendar application, it automatically records the input, execution, and forgetting of scheduled plans; it automatically determines the destination or movement route from GPS data; it analyzes the intensity of exercise or walking patterns from the accelerometer; and it uses the image recognition application to manage food in the refrigerator, identify medication packaging, and confirm garbage disposal status. Furthermore, the content entered by the user in the diary application or action records based on voice input are also acquired. Specifically, this Action Data Acquisition Unit performs inference processing using an edge AI chip on the user terminal. This Action Data Acquisition Unit inputs data to the edge AI model. This Action Data Acquisition Unit uses a 3-axis accelerometer time-series tensor and a GPS coordinate array as input data. The action data acquisition unit defines the dimension of the input data as a two-dimensional tensor composed of a time axis and a sensor axis. The elements of the input data include the terminal's tilt, changes in acceleration, and latitude and longitude information. As a preprocessing step, the action data acquisition unit performs gravity component removal and coordinate normalization. As a specific example of input data, the action data acquisition unit takes as input a tensor representing acceleration changes over the past 10 seconds or a coordinate array representing the movement trajectory. The action data acquisition unit obtains output data from the edge AI model. As output data, the action data acquisition unit outputs the probability distribution of action categories. The action data acquisition unit is configured such that each output value represents the probability of the user's current action (walking, running, stationary, moving by vehicle, etc.). As a specific example of output data, the action data acquisition unit outputs a walking probability of 0.9 or a stationary probability of 0.1. The action data acquisition unit utilizes this output result for subsequent processing. As a subsequent processing step, the action data acquisition unit performs a process that sends only the action tag and timestamp with the highest probability as lightweight metadata to the server. This operational data acquisition unit utilizes this technology to enhance privacy protection by not sending raw data to the server and improve processing speed at the edge. As a specific application area, this operational data acquisition unit is used as a foundational technology in privacy-conscious elderly care applications or behaviorally targeted advertising.

[0056] The Communication Data Acquisition Department uses user terminal call history, messaging application usage, SNS usage, voice recognition functions, video calling applications, email applications, camera applications, etc., to obtain data on the number of phone calls with family or friends, the number of messages sent and received, the frequency of SNS posts, the number of video calls, the number of emails sent, the number of photos or videos shared, the length of conversations, the diversity of topics, the presence or absence of emotional expression, the repetition of the same topics, the response speed to questions, the frequency of topic switching, the presence or absence of jokes or laughter, mispronouncing names, confusion in family structure, and confusion in dates or days of the week, etc., to determine the content or frequency of interpersonal communication. For example, the voice recognition function is used to transcribe the content of daily conversations, and the frequency or time period of communication, response speed, tendency to not reply to read messages, and the frequency of stickers or emoticons used are also obtained from the usage of SNS or messaging applications. Specifically, this Communication Data Acquisition Department installs a deep learning model that utilizes the attention mechanism in natural language processing. This Communication Data Acquisition Department inputs data into the deep learning model. This Communication Data Acquisition Department uses the word embedding matrix of the text as input data. This communication data acquisition unit defines the dimension of the input data as a two-dimensional matrix consisting of the product of the sequence length and the embedding dimension. The element content of the input data includes the transcribed text of the dialogue or word vectors of the message. As a preprocessing step, this unit performs sub-word lexicalization and positional encoding. A specific example of the input data is the lexical matrix of the sentence "I'm happy today" or the embedding matrix of a short post from an SNS. This unit obtains output data from a deep learning model. The output data is a context-sensitive feature vector. Each output value represents the semantic expression or sentiment polarity of the entire sentence. As a specific example of the output data, this unit outputs a high-dimensional vector representing positive sentiment or a latent vector representing the topic category. This unit utilizes the output for subsequent processing. As a subsequent processing step, this unit inputs the feature vector into a sentiment analysis classifier to calculate the final sentiment score. This communication data acquisition department utilizes this technology to improve the accuracy of context understanding and enhance robustness against noise. Specifically, this communication data acquisition department is used in automated response systems for customer support or negative public opinion detection systems.

[0057] The Data Integration Department integrates health data, activity data, and communication data for each user, generating an integrated dataset managed by time series. For example, it summarizes a day's vital signs, outing history, conversation content, social activity records, medication status, meal content, sleep duration, exercise volume, SNS usage, communication with family, and the implementation of hobbies into a single record, enabling the tracking of long-term changes or trends. The Data Integration Department also performs data correction for missing or outlier values, anonymization for privacy protection, and data normalization or standardization. Specifically, this Data Integration Department performs relational modeling using a graph database. The Data Integration Department inputs data into its graph construction algorithm. It uses feature nodes and edge information extracted from various data sources as input data. The dimensionality of this input data is defined as a node feature matrix and an adjacency matrix. The elements of the input data include nodes representing the user's health status, nodes representing their activity history, and weights representing the temporal and causal relationships between them. As preprocessing of the input data, the Data Integration Department performs scaling adjustments between heterogeneous nodes and weighted edge calculations. This data integration unit, as a specific example of input data, takes as input node vectors representing abnormal blood pressure or weight matrices connecting the edges of sleep deprivation and elevated blood pressure. This data integration unit obtains output data from a graph construction algorithm. As output data, this data integration unit outputs an integrated knowledge graph representation. This data integration unit is configured such that each output value signifies the user's overall state and the complex interactions between various factors. As a specific example of output data, this data integration unit outputs an optimized adjacency matrix or an overall graph embedding vector. This data integration unit utilizes this output result for subsequent processing. As subsequent processing, this data integration unit performs higher-order feature extraction by inputting the knowledge graph representation into a graph convolutional network. This data integration unit improves upon this technology, achieving a significant leap in the efficiency of complex relational computation and data retrieval speed. As a specific application area, this data integration unit is utilized in drug target exploration systems or factor analysis platforms for complex diseases.

[0058] The Cognitive Function Assessment Department analyzes and integrates datasets, leveraging generative AI to calculate cognitive function scores based on users' daily conversation content and behavior patterns. For example, it uses generative AI-based natural language processing technology to analyze conversation content in detail, incorporating indicators such as repetition of the same topic or pauses in speech, response speed to questions, vocabulary diversity, logical coherence, richness of emotional expression, use of jokes or metaphors, accurate recognition of the other person's name or date, consistency of conversation, frequency of topic switching, topic deviation, and expressions of care for the other person. Furthermore, changes in behavior patterns (such as a sharp decrease in the frequency of going out, reduced willingness to participate in activities of interest, increased forgetfulness, disrupted sleep rhythms, reduced mealtimes, increased forgetfulness of taking medication, and errors in household chores) are also reflected in the score. Finally, using generative AI, subtle changes in cognitive function are extracted from the user's overall life and quantified into a score out of 100. For generative AI, prompts are provided such as "Extract signs of cognitive function changes or decline from the user's dialogue and action records over the past month and score them" and "Detect patterns such as dialogue repetition, delayed response, and vocabulary reduction to assess the risk of cognitive function decline." Specifically, this cognitive function assessment unit implements a generative AI control architecture that combines prompt engineering and fine-tuning. This cognitive function assessment unit inputs data into the generative AI model. As input data, this cognitive function assessment unit uses prompt text vectors and the user's multimodal history tensor. This cognitive function assessment unit defines the dimension of this input data as a high-dimensional vector combining text tags and numerical features. The element content of the input data for this cognitive function assessment unit includes the embedded expression of the system-generated prompt text and statistical data of past dialogues or actions. As preprocessing of the input data, this cognitive function assessment unit performs variable embedding and tokenization processing into the prompt template. As a specific example of input data, this cognitive function assessment unit inputs the prompt vector of "Assess the risk of cognitive function decline from the dialogue embedding matrix of the past 30 days" or the actual dialogue history tensor. This cognitive function assessment unit obtains output data from a generative AI model. As output data, this cognitive function assessment unit outputs numerical data of cognitive function scores and text generation tags as the basis. This cognitive function assessment unit is configured such that each output value represents a quantitative assessment result, along with a logical explanation of the assessment. As a specific example of output data, this cognitive function assessment unit outputs a score of 75, or the text "The score decreased due to a reduction in the number of identified words." This cognitive function assessment unit utilizes this output result for subsequent processing. As subsequent processing, this cognitive function assessment unit performs the process of transmitting the score to the risk prediction unit and sending the basis text to the user interface unit for display on the screen. Through this technology, this cognitive function assessment unit improves and ensures the transparency of the assessment and reduces the illusion of AI.This cognitive function assessment unit is used as a specific application area in diagnostic support systems for healthcare professionals or in automated assessment platforms that require transparency.

[0059] The Risk Prediction Department uses a generative AI-based machine learning model to individually predict the risk of developing dementia, based on cognitive function scores and an integrated dataset. Multiple factors are inputted, including family history or past medical history, vital signs, changes in cognitive function scores, frequency of social activities, quality of communication, lifestyle habits, medication status, medical records, fall history, accident history, driving habits, economic status, living environment, and support systems. The risk level is calculated based on patterns learned by the generative AI from a vast amount of past data. The generative AI can also explain the risk factors, or which factors contribute to an increased risk, to the user or their family in natural language. The department provides prompts such as "Based on this user's integrated dataset, predict the risk of developing dementia and explain the risk factors" and "Extract lifestyle habits or behavior patterns that contribute to an increased risk and create improvement proposals." Specifically, this Risk Prediction Department is composed of a large-scale language model based on Transformer, which incorporates multilayer perceptrons and self-attention mechanisms. This risk prediction unit receives, as input data, a numerical vector containing time-series vital sign information and a natural language tag column containing textualized lifestyle information. The unit normalizes the numerical vector, converting it into a fixed-dimensional tensor. This unit then inputs the natural language tag column into a word embedding layer, generating a high-dimensional embedding vector. The unit combines these tensors and embedding vectors to generate, for example, a 512-dimensional integrated feature tensor. This integrated feature tensor is then input into the encoder of a large-scale language model. This large-scale language model is pre-trained using past electronic medical record data and lifestyle data. It is fine-tuned using a cross-entropy error function and gradient descent. From the decoder of the large-scale language model, the unit outputs a risk score with continuous values ​​ranging from 0.0 to 1.0. The unit further outputs an attention weight vector representing the contribution of features that contribute to an increase in risk. Based on this attention weight vector, the unit identifies the top three features with the highest contribution. These specific features are then input into the natural language generation module. This natural language generation module generates specific explanatory text such as "insufficient sleep time increases risk by 15%" and "insufficient exercise increases risk by 10%." When the output risk score exceeds a predetermined threshold, the system sends an emergency alert to the subsequent intervention plan generation department. Thus, this system quantitatively assesses the interactions between complex features in a high-dimensional space without relying on the subjective judgment of human doctors. As a result, this system achieves a significant leap in the accuracy of risk prediction. This technology can be applied to use cases such as preventative healthcare systems in facilities for the elderly or personalized health management applications.

[0060] Based on risk prediction results and the user's interests and lifestyle, the intervention plan generation department uses generative AI to automatically generate individually optimized prevention programs. When a user shows interest in activities such as puzzles, music, cooking, gardening, reading, travel, photography, handicrafts, Go, Shogi, karaoke, walking, dancing, yoga, volunteering, community activities, family communication, or pet ownership, the generative AI proposes various interventions tailored to their interests, including brain training programs, lifestyle improvement suggestions, introductions to social participation activities, and recommendations for specialized medical institutions. Using image generation technology based on generative AI, it also automatically generates user-specific brain training illustrations or quizzes, and personalized cognitive stimulation content. The generative AI is prompted with messages such as "Created a one-week brain training program and lifestyle improvement suggestions based on the user's interests and risks," "The proposal considers the user's interests or family structure, social participation activities, or communication promotion strategies," and "Generated memory training content using the user's photos or memories." Specifically, this intervention plan generation department is equipped with a computational module that incorporates reinforcement learning algorithms and generative adversarial networks. The intervention plan generation unit receives, as input data, a risk score output from the risk prediction unit and a one-hot vector representing a categorical variable indicating the user's interests and preferences. This unit combines this input data to generate a state vector. This state vector is then fed into a deep Q-network. The deep Q-network learns to use past user intervention history and subsequent changes in cognitive function scores as rewards. It updates the weight parameters to minimize the time difference error based on the Bellman equation. As the output of the deep Q-network, the intervention plan generation unit calculates a Q-value vector representing the action value for multiple intervention candidates. It selects the index with the maximum value in the Q-value vector to determine the optimal combination of intervention plans. If the determined intervention plan requires image content, the unit activates a generative adversarial network (GAN). This GAN consists of a generator and a discriminator. The generator combines feature vectors extracted from the user's past memories with a noise vector from the latent space as input. The generator repeats deconvolution operations to generate, for example, a 256-pixel square RGB image tensor. This discriminator compares the generated image tensor with the actual brain training image tensor to determine its authenticity. This generator advances learning by deceiving the discriminator, producing high-quality cognitive stimulus images optimized for the user. The intervention plan generation unit sends the generated image tensor and the textualized intervention plan to the subsequent AI coaching unit. Thus, this system departs from existing methods of providing a uniform procedure, achieving dynamic content generation that adapts to the user's individual situation. As a result, this system effectively increases user engagement and maximizes intervention effectiveness. This technology can be applied to use cases such as home care support systems or personalized education platforms for dementia prevention.

[0061] The AI-powered coaching department distributes generated prevention programs to user terminals, presenting proposals to users through interactive interfaces based on voice or text. Specific action proposals are given daily, such as "Let's do 10 minutes of puzzle brain training today," "Would you like to join a local community activity this afternoon?" "I've been going out less lately, how about a walk to refresh my mood?" "Let's have a video call with family today," "Would you like to try a new cooking recipe?" and "There's a local film screening this week." When users follow the proposals, their implementation is fed back to the server. The AI-powered coaching department uses generative AI to analyze the user's motivation or implementation, providing encouragement, reminders, or feedback at the optimal time and method. For the generative AI, prompts are provided such as "Generating the best encouragement or reminder text based on the user's activity implementation and emotional expression" and "Generating 10 examples of conversations to follow when the user fails to maintain activity." Specifically, this AI-powered coaching department has a multimodal dialogue engine that integrates a speech synthesis model and a natural language processing model. As input data, this AI-powered coaching department receives text-based action proposal data from the intervention plan generation department and voice waveform data acquired from the user's microphone. This AI tutoring unit performs a Fast Fourier Transform on the speech waveform data to extract an acoustic feature tensor containing Mel-frequency cepstral coefficients. This acoustic feature tensor is then input into a recurrent neural network (RNN). The RNN uses Long Short-Term Memory (LSTM) units to learn the dependencies of the time series. The RNN outputs an emotion vector representing the user's emotional state. This AI tutoring unit combines the emotion vector with action proposal data and inputs it into a prompt generation module. This prompt generation module uses a large-scale language model to generate encouraging text messages optimized for the user's current emotional state. The AI ​​tutoring unit inputs the generated text message into a speech synthesis model. This speech synthesis model converts the text message into a phoneme sequence, adjusts the pitch and intonation, and generates synthesized speech waveform data. The AI ​​tutoring unit outputs the synthesized speech waveform data to the user terminal's speaker. The AI ​​tutoring unit receives a binary flag from the user terminal indicating whether the user has implemented the proposed action. If the binary flag indicates no action has been taken, the AI ​​tutoring unit sends an alarm signal to the subsequent feedback management department. Therefore, this system is not simply a fixed-text notification, but rather a dynamic dialogue generation system implemented on the computer based on the user's real-time emotional state. As a result, this system effectively prevents users from becoming less motivated and enhances the naturalness of their interaction with the system. This technology can be applied to use cases such as smart speakers for the elderly or interactive health management robots.

[0062] The user interface (UI) unit provides various interfaces on the user terminal, including voice input / output, text display, image and video display, notifications, reminders, chatbots, calendar linkage, push notifications, vibration notifications, screen display, video playback, and music playback, enabling users to intuitively and continuously utilize the system. For example, it implements voice assistant-based dialogue, touch-based selection, camera-based image input, microphone-based voice input, reminder display in the notification bar, and automatic registration of scheduled events in the calendar. Specifically, this UI unit includes a graphical user interface control module and a sensor data integration module. As input data, this UI unit receives coordinate data from the user on the touchscreen, dynamic image tensors at 30 frames per second acquired from the camera, and audio stream data acquired from the microphone. This UI unit inputs the dynamic image tensors into a convolutional neural network. This convolutional neural network extracts landmark coordinates from the user's face and calculates the gaze direction vector. Based on the gaze direction vector, this UI unit determines which area on the screen the user is looking at. This UI unit performs rendering processing to automatically enlarge and display the icon of the gaze area. This UI unit inputs the audio stream data into a speech recognition engine and converts it into text data. This user interface unit uses an intent extraction algorithm to extract specific user commands from text data. Based on these commands, it generates and executes an SQL query to insert a scheduled record into the calendar database. Five minutes before the scheduled start time, it calls the operating system's notification API to generate a push notification. Simultaneously, it outputs a control signal to drive the terminal's vibration motor at a specific frequency. Thus, the system does not wait for explicit user input; it anticipates user intent from multimodal input such as sight or voice and dynamically changes the interface. As a result, the system achieves intuitive operation even for users unfamiliar with IT devices, reducing input error rates. This technology can be applied to use cases such as dedicated tablet terminals for the elderly or wearable devices that emphasize accessibility.

[0063] The Feedback Management Department centrally manages user activity implementation, cognitive function score trends, changes in risk levels, intervention plan implementation results, and feedback from family or medical institutions. It uses generative AI to provide optimal feedback or suggestions to users, their families, and healthcare professionals. For example, it generates feedback such as "Increased outings this week, and cognitive function scores are also trending upwards," "Recent conversations have become more repetitive, so increase communication with family," and "Please consider consulting a specialist." The generative AI provides prompts such as "Summarize the user's activity history and cognitive function score trends for one month, and create a report for the family," and "Summarize risk factors and recommend intervention strategies for medical institutions." Specifically, this Feedback Management Department has a time-series data parsing engine and a natural language generation model. As input data, it receives a time-series matrix representing activity implementation over the past 30 days and a floating-point array representing cognitive function score trends. The department then inputs the time-series matrix into an autoregressive integral moving average model to separate the trend and seasonal components of activity levels. This feedback management department inputs floating-point arrays into a gradient boosting decision tree algorithm to extract key features contributing to score variations. Combining these extracted features with trend components, the department generates a summary vector. This summary vector is then input into the encoder of a natural language generation model. This model employs a Transformer architecture and is pre-trained using a professional corpus for healthcare practitioners and a colloquial corpus for families. Metadata representing the output destination attribute is input into the decoder. When the output destination is a family, the department generates colloquial text data excluding professional terminology. When the output destination is a medical institution, the department generates structured report data containing specific numerical data and risk indicators. The generated text data and report data are then sent to subsequent external collaboration departments. Thus, this system eliminates the need for manual report creation from massive time-series data, automatically generating feedback in the optimal form based on the recipient's attributes. As a result, this system effectively prevents delays in information sharing and dramatically improves communication efficiency among relevant personnel. This technology can be applied to use cases such as remote medical monitoring systems or business support systems in nursing facilities.

[0064] The External Collaboration Department collaborates with medical institutions, nursing professionals, regional integrated support centers, families, social welfare workers, nursing managers, pharmacists, nutritionists, rehabilitation personnel, regional event organizers, interest groups, volunteer groups, and administrative agencies to securely share necessary information. For example, with the user's consent, it may send cognitive function scores or risk levels, and activity records to medical institutions; collaborate with nursing professionals to adjust care plans; or automatically acquire regional event information and propose it to users. For generative AI, it provides prompts such as "Create an information-sharing message for medical institutions or nursing professionals based on the user's health status and living conditions" and "Generate a message to regional event organizers conveying the user's willingness to participate." Specifically, this External Collaboration Department possesses a secure communication protocol stack and a data conversion gateway. As input data, this External Collaboration Department receives report data output from the Feedback Management Department, along with an encrypted digital token indicating the user's consent. This External Collaboration Department uses public-key cryptography to decrypt the digital token, verifying the information sharing permission level. If the permission level is deemed valid, this External Collaboration Department converts the report data into a JSON payload in the HL7 FHIR format, a standard specification for medical information exchange. This external collaboration department uses the TLS 1.3 protocol to encrypt the converted JSON payload and sends it to the medical institution's external server via a RESTful API. The department confirms successful data transmission by receiving an HTTP status code of 200. Furthermore, the department periodically retrieves XML data containing the event dates, times, and locations from RSS feeds or Web APIs provided by the organizers of regional events. The department parses the retrieved XML data and stores the event information in an internal database. It then inputs the event information into a natural language processing model to calculate the cosine similarity to the user's interest vector. Events with cosine similarity exceeding a specified threshold are selected and sent to the subsequent intervention plan generation department. Thus, this system automatically performs secure and seamless data collaboration on a computer with external systems that have different data formats. As a result, the system effectively eliminates data silos and builds a comprehensive support system that utilizes regional resources. This technology can be applied to use cases such as healthy life extension projects in smart cities or platforms for regional integrated care systems.

[0065] In this way, the system described in this embodiment, through close collaboration among its components, utilizes generative AI in a multi-layered and dynamic manner to comprehensively achieve dementia prevention activities optimized for each user and continuous support, real-time data analysis and feedback, collaboration with medical, nursing, family, and regional resources, emergency response, privacy protection, and improved user experience. Specifically, this system operates in a distributed computing environment combining multiple computing nodes via a high-speed network. The system containers each component as a microservice architecture, dynamically allocating resources using orchestration tools. The system receives streaming data continuously sent from user terminals as input data. The system temporarily stores the streaming data in an in-memory database and performs parsing without delay using a real-time processing engine. Based on the parsing results, the system executes inference processing of the generative AI model in parallel. The system configures intermediate tensors generated during inference processing in a shared memory space, transferring data at high speed between components. To protect user privacy, the system implements homomorphic encryption on all input data, executing machine learning model inference while maintaining encryption. The system only decrypts the inference results and sends them to the user terminal. Therefore, this system enables real-time processing of large-scale and complex multimodal data, which is impossible to achieve through existing single-server batch processing. As a result, the system effectively reduces the response time to changes in user status to the millisecond level, preventing delays in emergency situations. This technology can be applied to use cases such as national-scale healthcare platforms supporting millions of users simultaneously, or medical data analytics infrastructures requiring high security.

[0066] (Implementation steps) Step 1: Collection of User Information and Health Data. Users register in the system using smartphones, tablets, PCs, wearable devices, etc. The system automatically collects health-related data from health management applications or wearable devices, including heart rate, blood pressure, body temperature, steps, sleep time, weight, blood glucose levels, oxygen saturation, calories burned, resting heart rate, exercise intensity, wake-up and bedtime, past medical history, family medical history, medication history, allergy information, and regular physical examination results, and sends this data to the server. For example, it collects daily morning blood pressure measurements from a smartwatch, weekly step counts from an activity meter, and medication status from a medication management application, collecting diverse data in a time series. Specifically, in this step, the system executes a data collection agent in the background on the user's terminal. The system receives analog voltage signals from the photoplethysmography (PPG) sensor built into the wearable device as input data. The system inputs the analog voltage signals to an analog-to-digital converter, converting them into digital waveform data with a sampling rate of, for example, 100 Hz. The system applies a bandpass filter to the digital waveform data to remove baseline variations and high-frequency noise. This system uses a peak detection algorithm to calculate the heart rate interval from noise-removed waveform data, generating a numerical vector containing heart rate and heart rate variability indicators. The system further calls the healthcare API provided by the user terminal's operating system to obtain data such as step count or sleep time in JSON format. The system parses the obtained JSON data and combines it with the numerical vector to construct a health data payload. The system assigns a timestamp and user identifier to the health data payload and sends it asynchronously to the server using a message queue protocol. The system stores the received health data payload on the server side in a time-series database. Therefore, this system does not require explicit user input and continuously and completely collects heterogeneous mixed data from diverse sensors. As a result, this system improves the comprehensiveness and accuracy of data collection, ensuring the predictive accuracy of subsequent machine learning models. This technology can be applied to use cases such as remote monitoring systems for patients with chronic diseases or physical condition management systems for professional athletes.

[0067] Step 2: Collection of action data and communication data Using the user terminal's GPS function or accelerometer, calendar applications, activity tracking applications, voice input applications, image recognition applications, call logs, messaging applications, SNS usage, voice recognition functions, video call applications, email applications, camera applications, etc., the system acquires data on frequency of outings, distance traveled, amount of exercise, number of times participating in shopping or hobby activities, household chores performed, sleep patterns, meal times, bath times, number of toilet visits, TV viewing time, reading time, whether or not there is a nap, diversity of walking routes, number of calls, number of messages sent and received, frequency of SNS posts, length of conversations, diversity of topics, presence or absence of emotional expression, repetition of the same topic, response speed to questions, frequency of topic switching, etc., and sends this data to the server. Specifically, in this step, the system executes a context-aware module on the user terminal. The system receives latitude and longitude coordinate data obtained from the GPS module and 3-axis acceleration tensors obtained from the accelerometer as input data. The system inputs the coordinate data into a Kalman filter to correct positioning errors and generates a spatial vector representing the user's movement trajectory. This system performs a short-time Fourier transform on the acceleration tensor to extract a feature matrix in the frequency domain. This feature matrix is ​​then input into a support vector machine to classify action states such as walking, running, and stationary states. Further, the system utilizes the operating system's accessibility API to capture text data displayed on the messaging application screen. This acquired text data is input into a morphology parsing engine to generate feature vectors representing word frequency or part-of-speech distribution. Voice data from a phone call, captured from the microphone, is input into a speech interval detection algorithm to separate speaking intervals from silence intervals. The system accumulates the length of speaking intervals and calculates the call duration. This system integrates these action states, feature vectors, and call durations to construct an action-communication data payload. This payload is then sent to the server. Thus, unlike existing questionnaire surveys that rely on human self-reporting, this system automatically extracts objective and quantitative action indicators in the background. As a result, this system captures subtle changes in users' daily lives without omission, effectively contributing to the early detection of cognitive decline. This technology can be applied to use cases such as care services for elderly people living alone or employee mental health management systems.

[0068] Step 3: Data Integration and Preprocessing On the server, health data, activity data, and communication data are integrated for each user to generate an integrated dataset managed by time series. The system also performs correction for missing or outlier values, anonymization for privacy protection, and data normalization or standardization. For example, it summarizes a day's vital signs, travel history, conversation content, social activity records, medication status, meal content, sleep duration, exercise volume, SNS usage, communication with family, and activities related to hobbies into a single record. Specifically, in this step, the system executes a data pipeline processing engine built on the server. The system receives the health data payload and activity / communication data payload collected in steps 1 and 2 as input data. The system extracts timestamps from these payloads and groups the data into specified time windows (e.g., one-hour units). The system applies linear interpolation or spline interpolation algorithms to the grouped data to supplement missing values ​​caused by sensor communication errors, etc. The system uses an isolated forest algorithm as an outlier detection algorithm to identify and exclude or smooth records representing statistically abnormal values. This system inputs personally identifiable information such as a user's name or phone number into a hash function, converting it into an irreversible anonymized ID. It applies Z-score normalization or min-max scaling to numerical data, unifying the scale of all features. For textual data, it applies TF-IDF to convert it into a fixed-dimensional vector representation. Combining these preprocessed data, the system generates, for example, a multidimensional tensor arranging 1024-dimensional feature vectors along a time axis. This multidimensional tensor is then stored as an integrated dataset in a distributed file system. Thus, the system transforms unstructured data with varying sampling rates or formats into structured tensors that machine learning models can directly process. Consequently, this system improves the learning convergence speed of subsequent AI models and prevents computational errors during inference. This technology can be applied to use cases such as building large-scale medical data warehouses or generating learning data pipelines for multimodal AI.

[0069] Step 4: Cognitive Function Assessment (Generative AI Utilization) By analyzing and integrating datasets, generative AI is leveraged to calculate cognitive function scores from users' daily conversations or behavior patterns. The generative AI is fed information such as, "Extract signs of changes or decline in cognitive function from the user's conversation and behavior records over the past month, and assign a score." The system uses prompts such as "Detect patterns of dialogue repetition, delayed responses, and vocabulary reduction to assess the risk of cognitive decline." Generative AI analyzes the dialogue content in detail using natural language processing (NLP) technology, using evaluation indicators such as repetition of the same topic or pauses, response speed to questions, vocabulary diversity, logical connections, richness of emotional expression, use of jokes or metaphors, accurate identification of the other party's name or date, dialogue consistency, frequency of topic switching, topic deviation, and expressions of care for the other party. Changes in behavior patterns are also reflected in the scoring. Specifically, in this step, the system possesses a large-scale language model specialized for NLP and a graph neural network for time series analysis. The system receives text vectors and action trajectory vectors contained in the multidimensional tensor generated in step 3 as input data. The system inputs the text vectors into the encoder of the large-scale language model. This large-scale language model has hundreds of millions of parameters and has been pre-trained through an autoregressive language modeling task. Within the internal attention layer of the large-scale language model, the system calculates word dependencies, entropy values ​​representing vocabulary diversity, and coherence scores representing logical connections. This system inputs action trajectory vectors into a graph neural network. This graph neural network constructs a graph structure with user access locations as nodes and movement paths as edges, calculates the degree centrality or clustering coefficients of the nodes, and generates graph embedding vectors representing the diversity of actions. The system inputs entropy values, coherence scores, and graph embedding vectors into a fully connected layer. This fully connected layer uses the ReLU activation function for non-linear transformation, outputting the final cognitive function score as a continuous value from 0 to 100. The system stores the output cognitive function score in a time-series database and calculates the difference with past scores. If the difference exceeds a predetermined threshold, the system sends a trigger signal to the subsequent risk prediction department. Thus, this system replaces existing paper-and-pencil tests for cognitive function assessment, continuously and non-invasively quantifying cognitive function from daily digital footprints. As a result, this system achieves the technical effect of guaranteeing the objectivity of the assessment and eliminating biases caused by the examiner's subjectivity. This technology can be applied to use cases such as digital biomarker development platforms or dementia screening applications.

[0070] Step 5: Risk Prediction (Utilizing Generative AI) Based on cognitive function scores and an integrated dataset, a generative AI-based machine learning model is used to individually predict the risk of developing dementia. The generative AI is provided with prompts such as "Predict the risk of developing dementia based on the user's integrated dataset and explain the risk factors" and "Extract lifestyle habits or behavior patterns that contribute to an increased risk and create improvement proposals." Multiple factors are used as input, including family history or past medical history, vital signs, changes in cognitive function scores, frequency of social activities, quality of communication, lifestyle habits, medication status, medical records, fall history, accident history, driving habits, economic status, living environment, and support systems. The risk level is calculated, and the explanation of risk factors is also output by the generative AI in natural language. Specifically, in this step, the system possesses an ensemble learning model and a causal inference engine. The system receives the time-series array of cognitive function scores calculated in step 4 and the multidimensional tensor generated in step 3 as input data. The system inputs the time-series array into a long short-term memory network to extract hidden state vectors representing the downward trend of the scores. This system inputs a multidimensional tensor into a random forest algorithm to calculate the importance score of each feature. Combining the hidden state vector and importance scores, the system generates an integrated feature vector. This integrated feature vector is then input into a causal inference engine. This engine uses a Bayesian network to model the causal relationships between features, calculating the probabilistic impact of specific lifestyle habits on future cognitive decline. As the output of the causal inference engine, this system generates probability distribution vectors representing the probability of developing dementia in 1, 3, and 5 years. Based on the highest probability value in the probability distribution vectors, the system categorizes the risk level as "low," "medium," or "high." The system then inputs the categorized risk levels and the causal path obtained from the causal inference engine into a natural language generation model. This model generates specific risk factor descriptions, such as "Reduced social activity over the past 3 months increases the risk of developing dementia in 3 years by 20%." Finally, the system sends the generated text and risk levels to the subsequent intervention plan generation department. Therefore, this system goes beyond simply indicating correlations; it unravels complex causal relationships in high-dimensional data spaces and outputs predictions based on these relationships. As a result, this system effectively improves the transparency and interpretability of predictions, enhancing user confidence. This technology can be applied to use cases such as risk assessment systems in insurance companies or personal health record systems for preventative medicine.

[0071] Step 6: Generation of individually optimized intervention plans (utilization of generative AI) Based on risk prediction results and users' interests and lifestyles, generative AI automatically generates individually optimized prevention programs. The generative AI is provided with prompts such as "Create a one-week brain training program and lifestyle improvement suggestions tailored to the user's interests and risks," "Proposals consider social participation activities or communication promotion strategies based on the user's hobbies or family structure," and "Generate memory training content using the user's photos or memories." Multiple interventions, including brain training programs, lifestyle improvement suggestions, introductions to social participation activities, and recommendations for specialized institutions, are combined to create proposals. Image generation technology also automatically generates user-specific brain training illustrations or tests, and personalized cognitive stimulation content. Specifically, in this step, the system has a diffusion model for content generation and a collaborative filtering algorithm for the recommendation system. As input data, the system receives risk level and risk factor description text generated in step 5, as well as interest vectors obtained from the user's profile database. The system calculates the inner product between the interest vectors and other users' vectors, extracting a list of IDs of user groups with high similarity who have previously achieved high-effective intervention programs. The system combines the extracted ID list and risk factor description text and inputs it into the prompt encoder. This prompt encoder converts textual information into conditional vectors in the latent space. The system then inputs these conditional vectors into a diffusion model. Starting from Gaussian noise, this model generates high-quality brain training image tensors related to the user's preferences through repeated denoising processes using the U-Net architecture (approaching dozens of steps). Simultaneously, the system inputs the conditional vectors into a Transformer model for text generation, producing question and answer texts for image-related quizzes. The system then packages the generated image tensors, question texts, and answer texts into a JSON-formatted intervention plan object. This intervention plan object is then sent to the subsequent AI tutoring department. Thus, this system moves away from the existing method of distributing pre-prepared static content, dynamically generating content perfectly aligned with the user's current state and preferences. As a result, this system effectively prevents user boredom and dramatically increases engagement in the intervention process. This technology can be applied to use cases such as adaptive learning systems or rehabilitation support systems in the education field.

[0072] Step 7: Proposal development, implementation, and monitoring through AI guidance (generative AI application) The generated prevention program is distributed to user terminals, where it is presented to users through an interactive interface based on voice or text. The generative AI is provided with tools to "generate optimal encouraging or reminding text based on the user's activity and emotional expression" and "generate 10 examples of conversations to be made when a user fails to maintain activity." The prompt text reads: "Let's do 10 minutes of jigsaw puzzle brain training today." Every morning, the prompt reads: "Let's do 10 minutes of jigsaw puzzle brain training today." Specific action proposals such as "Would you like to join a nearby community activity this afternoon?", "I haven't been going out much lately, how about going for a walk to refresh my mood?", and "Let's have a video call with family today?" are monitored to track user reactions and implementation, dynamically optimizing feedback or reminders for activity continuation or motivation maintenance. Specifically, in this step, the system features a reinforcement learning-based dialogue management module and a convolutional neural network for emotion recognition. The system receives the intervention plan object generated in step 6 as input data, along with facial image tensors acquired from the user's front-facing camera. The system inputs the facial image tensors into the convolutional neural network to extract feature maps representing facial muscle movements and calculate probability vectors representing emotional states such as joy, sadness, and fatigue. The system inputs the probability vectors and the intervention plan object into the dialogue management module. This module evaluates the current state based on a Markov decision process, determining the optimal timing and speaking style for proposing interventions to the user. Based on the determined speaking style, the system uses a large-scale language model to generate personalized proposal texts such as "You seem a little tired, would you like to take a simple 5-minute quiz?". This system uses a speech synthesis engine to convert proposal text into speech waveform data, which is then played back through the user's speaker. The system collects performance data in real time, such as screen click intervals and accuracy rates, during the user's intervention program implementation. If performance data falls below a specified benchmark, the system feeds back control signals to the intervention plan generation department to reduce the difficulty. Upon completion of the intervention, the system generates and displays reward text to enhance the user's sense of accomplishment. Therefore, this system is not a one-way notification but rather infers the user's emotional state or fatigue level in real time, achieving two-way adaptive interaction on the computer. As a result, the system effectively reduces the user's psychological burden and promotes long-term behavioral change. This technology can be applied to use cases such as virtual coaches in fitness applications or digital therapies supporting medication adherence.

[0073] Step 8: Regular reassessment and feedback (generative AI utilization) Cognitive function assessments or risk predictions are periodically re-implemented, and generative AI is used to analyze changes in scores or behavior patterns. The generative AI is provided with prompts such as "summarize the user's activity history and cognitive function score trends over the past month, and create a report text for the family" and "summarize risk factors and recommend intervention strategies for medical institutions." In cases of a sharp decline in cognitive function scores, or when abnormal patterns are detected in conversation content, the generative AI automatically summarizes the content and sends alerts to the user or family, such as "recommendation to a specialized institution." Specifically, in this step, the system possesses a distributed data processing framework for batch processing and anomaly detection algorithms. The system receives time-series data of cognitive function scores accumulated over the past month, as well as records from a structured database containing the user's activity logs, as input data. The system inputs the time-series data into an autoregressive moving average model to analyze score trends and periodicity. The system calculates the residual between the expected value predicted from the analysis results and the actual score, marking data points with residuals exceeding statistically permissible limits as outliers. When outliers are detected, the system inputs the activity logs for that period into a natural language processing model. This natural language processing model extracts specific patterns from activity logs, such as a sharp decrease in the number of outings or an extreme reduction in vocabulary during conversations. The system inputs the extracted patterns and score trends into a Transformer model for summary generation. This Transformer model uses an encoder-decoder architecture to convert the extracted features into report text in human-readable natural language. When the recipient of the report text is a family member, the system performs filtering to replace technical terms with colloquial expressions. When the recipient is a medical institution, the system generates a structured report with added numerical data such as p-values ​​or confidence intervals indicating the statistical significance of outliers. The system pushes the generated report text and the structured report to the designated recipient's terminal via a secure communication channel. Thus, this system automatically extracts subtle changes that are easily overlooked by humans from massive log data and issues alerts at appropriate times. As a result, this system can provide early intervention for sharp declines in cognitive function and prevent the progression to severe illness. This technology can be applied to use cases such as resident management systems in elderly care facilities or triage systems in telemedicine.

[0074] Step 9: External Collaboration and Emergency Response With the user's consent, cognitive function scores or risk levels, activity history, etc., will be shared securely with medical institutions, care providers, regional integrated support centers, families, social welfare workers, care managers, pharmacists, nutritionists, rehabilitation personnel, regional event organizers, etc.

[0075] For example, cognitive function scores or risk levels can be sent to healthcare institutions, care plans can be adjusted in collaboration with care providers, regional activity information can be automatically retrieved and proposed to users, and families or healthcare institutions can be automatically notified in emergencies such as falls or sudden cognitive decline. Specifically, in this step, the system has an event-driven message broker and an API gateway. The system receives the structured report generated in step 8 as input data, as well as impact waveform data representing sudden changes in gravity obtained from the user's terminal's accelerometer. The system inputs the impact waveform data into a support vector machine to calculate the probability of a fall. If the probability of a fall exceeds a predetermined threshold (e.g., 95%), the system immediately initiates emergency interruption processing. In emergency interruption processing, the system packages GPS coordinates representing the user's current location and a tensor of recent vital signs information and sends an HTTPS POST request to the emergency medical services server. The system also uses the WebSocket protocol to push and distribute real-time alert notifications to family smartphones. In normal use outside of emergencies, the system converts the structured report into XML format conforming to the HL7 FHIR standard. This system periodically sends the converted XML data to the RESTful API endpoint of the electronic medical record system used by healthcare institutions or nursing operators. The system further parses activity information retrieved from a regional activity database using a natural language processing model and matches it with the user's interest vector. The system generates URLs for participating in activities with high matching scores and displays them on the user's terminal interface. Thus, this system does not remain confined to its internal system but rather builds an autonomous network that seamlessly integrates with external healthcare and nursing infrastructure. As a result, this system achieves the technical effect of minimizing emergency response time and significantly improving user safety. This technology can be applied to use cases such as nursing networks in smart homes or autonomously-led regional integrated nursing platforms.

[0076] (specific use case) For example, when a couple around 80 years old uses the system of this invention, the health data acquisition unit first automatically acquires health-related data such as heart rate, blood pressure, steps, sleep time, weight, blood sugar levels, medication history, past medical history, family medical history, and regular physical examination results from wearable devices such as smartwatches, activity meters, blood pressure monitors, and thermometers, as well as health management applications, and sends it to the server. The activity data acquisition unit uses the user terminal's GPS or accelerometer sensor, calendar applications, activity tracking applications, voice input applications, image recognition applications, etc., to acquire data such as frequency of outings, distance traveled, amount of exercise, number of times participating in hobby activities, household chores performed, sleep rhythm, meal times, bathing times, TV viewing time, reading time, and the diversity of walking routes. The communication data acquisition unit uses call history, messaging applications, SNS usage, voice recognition functions, video call applications, email applications, etc., to acquire data such as the number of phone calls with family or friends, number of messages sent and received, frequency of SNS posts, length of conversations, diversity of topics, presence or absence of emotional expression, repetition of the same topics, response speed to questions, and frequency of topic switching. Specifically, this system possesses a relational database that associates and manages multiple user accounts, as well as edge computing devices. As input data, the system receives 50 samples per second of 3-axis acceleration tensor data from the husband's smartwatch, and systolic and diastolic blood pressure data transmitted via Bluetooth from the wife's blood pressure monitor. The system executes a lightweight convolutional neural network on the edge computing device to infer probability vectors for action categories such as walking, housework, and sleep in real time from the acceleration tensor. The system serializes the inferred action category probability vectors and blood pressure data, along with timestamps and user IDs, into a JSON payload. The system sends this JSON payload to a cloud server using the MQTT protocol. On the cloud server side, the system receives the couple's conversational voice data from the microphone of a smart speaker. The system inputs the conversational voice data into a speaker logging algorithm to separate the husband's and wife's speaking intervals. The system inputs the separated speaking intervals into a speech recognition model, converting them into text data. The system inputs the text data into a natural language processing model to calculate the number of turns, the frequency of overlap, and the sentiment polarity score. This system extracts these communication indicators as numerical vectors. It integrates health data, activity data, and communication data to construct multidimensional feature tensors for each couple. Therefore, this system does not rely on a single device, enabling the synchronization and processing of multimodal data asynchronously generated from multiple IoT devices. As a result, this system effectively quantifies the complex interactions in the daily lives of the elderly and improves the resolution of data collection. This technology can be applied to use cases such as family care systems in smart homes or facility-oriented solutions that simultaneously manage the health status of multiple individuals.

[0077] The Data Integration Department integrates health data, action data, and communication data for each user, generating an integrated dataset managed by time series. The Cognitive Function Assessment Department analyzes this integrated dataset, utilizing generative AI to calculate cognitive function scores from users' daily conversation content or action patterns. For example, it provides prompts to the generative AI such as "Extract signs of changes or decline in cognitive function from the user's conversation and action records over the past month and score them," and "Detect patterns such as repetition in conversation, delayed responses, and reduced vocabulary to assess the risk of cognitive function decline." The generative AI analyzes the conversation content in detail, using indicators such as repetition of the same topic or hesitation, response speed to questions, vocabulary diversity, logical connections, richness of emotional expression, use of jokes or metaphors, accurate identification of the other party's name or date, consistency of conversation, frequency of topic switching, topic deviation, and expressions of care for the other party. Changes in action patterns are also reflected in the score. Specifically, this Data Integration Department uses an in-memory data grid to perform high-speed data combination processing. As input data, this Data Integration Department receives multiple unstructured data streams containing health data, action data, and communication data. This data integration unit maps these data streams onto a common time axis and fills in missing values ​​using spline interpolation, thereby generating, for example, a 2048-dimensional integrated feature tensor. This cognitive function assessment unit inputs this integrated feature tensor into a large-scale language model and a time-series Transformer model. The time-series Transformer model detects periodic or anomalous deviations in action patterns and outputs an action anomaly score vector. This large-scale language model uses a self-attention mechanism to parse the dialogue text data. As input, this large-scale language model receives the frequency of occurrence of the same noun within the text data, metadata representing the time interval between utterances, and syntactic tree data representing grammatical complexity. Based on these inputs, this large-scale language model calculates a repetition score representing topic repetition, a fluency score representing speech impediments, and a TTR value representing lexical diversity. This cognitive function assessment unit inputs these scores and the action anomaly score vector into a fully connected neural network. This fully connected neural network linearly combines each indicator by optimal weights, applies a sigmoid function, and outputs a final cognitive function score normalized to the range of 0 to 100. This cognitive function assessment unit then sends the calculated score to the subsequent risk prediction unit. Therefore, this system replaces the qualitative assessment conducted by human experts through lengthy interviews with quantitative calculations of high-dimensional data performed by computers. As a result, the system effectively ensures the objectivity and reproducibility of the assessment and enables the early detection of even minor cognitive decline. This technology can be applied to use cases such as automated consultation systems in telemedicine or as tools for calculating objective assessment indicators in clinical trials.

[0078] The Risk Prediction Department uses a generative AI-based machine learning model to individually predict the risk of developing dementia, based on cognitive function scores and an integrated dataset. For example, the generative AI is given prompts such as "Predict the risk of developing dementia based on this user's integrated dataset and explain the risk factors" and "Extract lifestyle habits or behavior patterns that contribute to an increased risk and create improvement proposals." Multiple factors are taken as input, including family history or past medical history, vital signs, changes in cognitive function scores, frequency of social activities, quality of communication, lifestyle habits, medication status, medical records, fall history, accident history, driving habits, economic status, living environment, and support systems. The risk level is calculated, and the explanation of risk factors is also output by the generative AI in natural language. Specifically, this Risk Prediction Department includes a hybrid inference engine that integrates a gradient boosting decision tree algorithm and a Transformer model for natural language generation. As input data, the Risk Prediction Department receives a time-series array of cognitive function scores and a feature vector consisting of dozens of categorical and continuous variables containing family history or vital signs information. This risk prediction unit inputs feature vectors into a gradient boosting decision tree algorithm. This algorithm, by learning a majority of decision trees in a concatenated manner, captures the complex nonlinear interactions between features and outputs a continuous score from 0.0 to 1.0 representing the future probability of developing dementia. This risk prediction unit further executes the SHAP algorithm to calculate the contribution of each feature to the output score. This unit extracts the top 5 features with the highest contributions (e.g., reduced sleep duration, forgetting to take specific antihypertensive medication, reduced frequency of outings, etc.). This unit inputs structured data representing the extracted features and their contributions into the encoder of a Transformer model. The Transformer model outputs text data from the decoder, such as "The recent decrease in outing frequency increases the risk by 15%. Furthermore, sleep disturbances also have an impact." This unit performs a threshold determination on the continuous score to determine the risk level and sends it along with the text data to the subsequent intervention plan generation unit. Thus, this system provides mathematically interpretable results for the predictions of machine learning models, which are often considered black boxes. As a result, this system powerfully supports healthcare professionals or users in trusting predictive outcomes and making decisions about next steps. This technology can be applied to use cases such as risk assessment tools in preventative medicine clinics or healthcare guidance support systems for health insurance portfolios.

[0079] The intervention plan generation department uses generative AI to automatically generate individually optimized prevention programs based on risk prediction results and the user's interests and lifestyle. For example, it provides prompts such as "Create a one-week brain training program and lifestyle improvement suggestions tailored to the user's interests and risks," "Proposals consider the user's hobbies or family structure, including social participation activities or communication promotion strategies," and "Generate memory training content using the user's photos or memories." It combines various interventions, including brain training programs, lifestyle improvement suggestions, introductions to social participation activities, and recommendations for specialized medical institutions, to propose solutions. It also automatically generates user-specific brain training illustrations or tests, and personalized cognitive stimulation content using image generation technology. Specifically, the intervention plan generation department includes a constrained optimization solver and a multimodal generative AI model. As input data, the intervention plan generation department receives risk levels output from the risk prediction department, a profile vector representing the user's hobbies, and a matrix representing past intervention history. This input data is then fed into the constrained optimization solver. This solver uses user fatigue or available time as constraints to calculate the optimal allocation ratio of intervention categories (e.g., exercise, cognitive training, social participation) that maximizes the expected value of the intervention effect. Based on the calculated allocation ratio, the intervention plan generation unit inputs cue vectors into a multimodal generative AI model. This multimodal generative AI model has an architecture that integrates a large language model for text generation and a latent diffusion model for image generation. The multimodal generative AI model receives image tensors of the user's past travel photos as input, adds noise to the latent space, and conditionally denoises, thereby generating a processed image tensor for spotting errors in a portion of the photo, along with the text "Where was this photo taken?". Further, assuming the user's hobby is "gardening," the intervention plan generation unit retrieves activity information from local gardening circles via web scraping and generates guiding text encouraging participation. The intervention plan generation unit packages the generated image tensor and text data as a weekly schedule into a JSON-formatted intervention plan object. This intervention plan object is then sent to the subsequent user interface unit. Therefore, this system does not impose a uniform program, but rather dynamically generates content deeply rooted in the user's personal context. As a result, the system effectively stimulates users' spontaneous participation and dramatically increases the persistence rate of intervention programs. This technology can be applied to use cases such as personalized digital cognitive behavioral therapy applications or entertaining rehabilitation systems for the elderly.

[0080] The AI ​​tutoring department distributes the generated prevention programs to user terminals, proposing them to users through voice- or text-based interactive interfaces. For example, it provides the generative AI with prompts such as "Generate the best encouragement or reminder based on the user's activity performance and emotional expression" or "Generate 10 examples of conversations to make when a user fails to maintain activity." Every morning, it provides specific action proposals such as "Do 10 minutes of puzzle brain training today," "Would you like to join a nearby community activity this afternoon?" "I've been going out less lately, how about going for a walk to refresh my mood?" or "Have a video call with your family today," monitoring the user's response or implementation and dynamically optimizing feedback or reminders for sustained activity or maintaining motivation. Specifically, this AI tutoring department includes a natural language generation engine based on a large language model. This engine performs dynamic prompt generation processing based on the user's context. The input data to this engine consists of a multi-dimensional vector representing the user's activity performance and a numerical score representing emotional expression. Specific examples of activity performance include a 7-dimensional integer array representing the number of steps taken in the past 7 days, or a scalar value representing the number of times the puzzle game was performed. As a specific example of emotional expression, the input data includes a floating-point vector ranging from 0.0 to 1.0, representing joy or sadness, output from the speech parsing module. This engine normalizes this input data as preprocessing and embeds it into a text-formatted prompt template. This prompt is then fed into a large language model. The output data from this engine consists of a generated text string and probability distribution scores representing the confidence level of each action proposal. Specific examples of the output text include natural language articles encouraging the user or reminders prompting specific activities. Specific examples of probability distribution scores include numerical arrays such as a score of 0.85 for a proposal walk and 0.42 for a proposal puzzle. As a subsequent processing step, the AI ​​tutor compares the output probability distribution scores with a pre-set threshold. The AI ​​tutor only extracts proposals with scores exceeding the threshold and sends them to the display control module on the user's terminal. The display control module displays the received text on the screen. Alternatively, the display control module activates the speech synthesis engine to read the text aloud. Through this processing, the system selects the most suitable intervention for the user's current state in real time. This system prevents user fatigue caused by irrelevant notifications. This system effectively prompts users to change their behavior and dramatically improves the persistence rate of prevention programs. It proves highly useful in specific application areas such as dementia prevention applications for the elderly and rehabilitation support systems.

[0081] Cognitive function assessments or risk predictions are periodically re-implemented, using generative AI to analyze changes in scores or behavior patterns. For example, the generative AI is provided with prompts such as "Summarize the user's activity history and cognitive function score trends over one month to create a report for the family" or "Summarize risk factors and recommend intervention strategies for medical institutions." In cases of a sharp decline in cognitive function scores or the detection of abnormal patterns in conversations, the generative AI automatically summarizes the content and sends alerts to the user or family, such as "Recommendation to a specialized institution." Specifically, this system periodically performs a reassessment process using a time-series data analysis model. This analysis model employs a neural network architecture with a self-attention mechanism. The input data to this analysis model is a time-series tensor representing the activity history and cognitive function score trends over a certain period. Specific examples of activity history include a 30-dimensional integer array representing daily steps or a 30-dimensional floating-point array representing sleep time. Specific examples of cognitive function score trends include a 4-dimensional floating-point array representing scores calculated weekly. This analysis model combines these input data along the time axis and performs normalization processing. The output data of this analytical model includes anomaly detection flags, summary text, and importance scores for risk factors. Specific examples of anomaly detection flags include binary values ​​of 0 (normal) and 1 (abnormal). Specific examples of summary text include natural language strings describing the factors contributing to the score decline. Specific examples of importance scores include numerical vectors such as a contribution of 0.6 for insufficient sleep and 0.3 for insufficient exercise. As a subsequent processing step, this system determines whether the anomaly detection flag is 1. If the flag is 1, the system immediately activates the alarm generation module. This alarm generation module sends a push notification containing the generated summary text to the family's terminals. Through this processing, the system mathematically detects subtle score decline trends that are easily missed by human visual inspection. This system effectively achieves early anomaly detection and prevents delays in medical intervention. In specific applicable fields such as remote medical monitoring and home care services, this system makes a significant contribution to ensuring user safety.

[0082] Thus, the system of this invention comprehensively collects and integrates health data, action data, and communication data from multiple perspectives. It uses generative AI to achieve cognitive function assessment, risk prediction, generation of individually optimized intervention plans, AI tutoring, and feedback generation, continuously and effectively providing dementia prevention activities optimized for each user. Specifically, this system constructs a high-level computational pipeline for integrating and processing multimodal data. This pipeline includes a group of encoders that map heterogeneous data to the same feature space. The input data to this pipeline is an integrated tensor combining health data, action data, and communication data. Specific examples of health data include time-series arrays of heart rate. Specific examples of action data include trajectory data with GPS coordinates. Specific examples of communication data include speech waveform data of conversations. This pipeline performs preprocessing on these input data, including missing value imputation and dimensionality reduction. The output data from this pipeline is a parameter set defining the optimized intervention plan. Specific examples of the parameter set include numerical values ​​representing the intensity of the exercise task, timestamps representing the task execution time, and identifiers of recommended communication partners. As subsequent processing, this parameter set is applied to the overall system's motion control module. This motion control module updates the action schedules of each subsystem based on the parameter set. This system, through its processing, provides a data-driven preventative approach fundamentally different from existing rule-based systems that rely on a single data source. It leverages the technology to dynamically adjust interventions based on subtle changes in the user's state, maximizing preventative effectiveness. This system offers innovative solutions in specific application areas such as personalized medicine and next-generation healthcare platforms.

[0083] (Application Example 1) The specific processing flow corresponding to Use Case 1 will be described below. The various parts of the system described below are implemented by the data processing device 12 and the intelligent device 14. Furthermore, the data processing device 12 is referred to as the "server" and the intelligent device 14 is referred to as the "terminal".

[0084] (Implementation methods for carrying out the present invention) The embodiments used to carry out the present invention will be described in a more detailed and comprehensive manner to further enhance the specificity.

[0085] This invention relates to a system that continuously and in real-time collects various data from each user in nursing facilities or home care settings, including health status, daily activities, communication patterns, and care records. It utilizes generative AI to accurately assess changes in cognitive or physical function, predict risks, and automates and enhances individualized optimized care plans or rehabilitation menus, life support proposals, feedback to family or caregivers, early anomaly detection, and medical collaboration. Specifically, this system employs a distributed processing architecture that enables edge computing devices to collaborate with cloud servers. The system includes an event-driven data collection module that processes streaming data asynchronously from various data sources. Input data to this module consists of streams of raw data sent from various sensors. Specific examples of raw data include time-series heart rate data sampled at 1 Hz or arrays of binary states output from indoor dynamic monitoring sensors. This module timestamps these input data and applies noise reduction filters. Output data from the system consists of structured data constituting risk prediction scores and care plans. Specific examples of risk prediction scores include floating-point numbers representing fall risk in the range of 0.0 to 1.0. As a concrete example of structured data, this includes records in JSON format containing task names, execution times, and responsible person IDs. As a subsequent processing step, the system activates the task allocation engine when the risk prediction score exceeds a predetermined threshold. Based on the structured data, this engine sends task execution instructions to the terminal of the best caregiver. Through this processing, the system parses massive amounts of data in real time, deriving optimal response strategies at speeds exceeding human cognitive capabilities. This system significantly reduces the workload of caregivers and standardizes the quality of care. In specific applicable areas such as large specialized nursing homes or those providing wide-area access care services, this system achieves a leap in operational efficiency.

[0086] First, the health data acquisition department automatically acquires various health-related data from wearable devices such as smartwatches, activity meters, blood pressure monitors, thermometers, pulse oximeters, and blood glucose meters, as well as from health management applications, electronic medical records from regular physical examinations, and medication management applications installed on smartphones or tablets. This data includes heart rate, blood pressure, body temperature, steps, sleep duration, weight, blood glucose levels, oxygen saturation, calories burned, resting heart rate, exercise intensity, wake-up and bedtime times, past medical history, family medical history, medication history, allergy information, and results of regular physical examinations. For example, it can acquire daily morning blood pressure measurements from smartwatches via Bluetooth, weekly step counts from activity meters, monthly weight changes from scales via Wi-Fi, medication status from medication management applications, and can also automatically link to electronic medical records from medical institutions, including blood test results, diagnostic images, and doctor's opinions. Furthermore, when a user visits multiple medical institutions, the system also integrates data from these institutions, automatically detecting and correcting duplicate or contradictory data. Specifically, this health data acquisition unit includes a data cleaning and normalization engine for integrating heterogeneous data formats. The input data to this engine is unstructured and semi-structured data obtained from multiple external systems. Specific examples of input data include blood pressure data described in JSON format, step count data described in CSV format, and electronic medical record data described in HL7 format. This engine performs pattern mapping on these input data, unifying units or interpolating missing values. The output data from this engine is a normalized health data vector converted into a format that can be processed by machine learning models. Specific examples of output data include a 128-dimensional numerical vector with each vital sign as an element, or a one-hot encoded vector representing past medical history. As subsequent processing, this health data acquisition unit stores the generated normalized health data vector in a time-series database. Simultaneously, this health data acquisition unit inputs this vector into an outlier detection algorithm to determine if statistically significant outliers exist. This system, through this processing method, absorbs noise caused by differences in data format or input errors, resulting in a significant improvement in the inference accuracy of subsequent AI analysis. In specific application areas such as regional integrated nursing systems that connect multiple medical institutions or nursing professionals, this system provides a highly reliable data foundation.

[0087] The Daily Activity Data Acquisition Department uses the user's terminal's GPS function or accelerometer, gyroscope, calendar application, activity tracking application, voice input application, image recognition application, etc., to acquire daily activity patterns such as frequency of outings, distance traveled, amount of exercise, number of times shopping or participating in hobbies, completion of housework, sleep rhythm, meal times, bathing times, number of toilet visits, TV viewing time, reading time, whether or not there is a nap, diversity of walking routes, and use of transportation. For example, it links with the calendar application to automatically record the input, execution, and forgetting of scheduled plans; it automatically determines the destination or movement route from GPS data; it analyzes the intensity of exercise or walking patterns from the accelerometer; and it uses the image recognition application to manage food in the refrigerator, identify medication packaging, and confirm garbage disposal status. Furthermore, it also collects content entered by the user into the diary application or activity records entered via voice input, such as subjective activity records like "I took a walk in the nearby park today," "I watched TV for 2 hours after lunch," and "I prepared dinner myself," as data. In addition, data such as the number of times a person enters or leaves a room, the number of times they use the toilet at night, the number of times they get out of bed, and changes in room temperature or humidity are also collected from IoT devices such as door opening / closing sensors, bed sensors, lighting sensors, and temperature and humidity sensors within the facility to detect signs of daily routines or unusual behavior. Specifically, this daily activity data acquisition unit includes an action analysis module that combines sensor fusion technology and time-series pattern recognition algorithms. The input data to this module is a multi-dimensional time-series data stream acquired synchronously from multiple sensors. Specific examples of the input data include a time-series array representing latitude and longitude GPS coordinates, continuous tensor data obtained from a 3-axis accelerometer, and a binary state array representing the opening and closing status of doors from IoT sensors. This module applies a Kalman filter to these input data to remove noise and performs feature extraction. The output data from this module is the inferred action label and its confidence score. Specific examples of action labels include categorization strings such as sleep, walking, and eating. Specific examples of confidence scores include values ​​such as a probability of 0.92 for walking and a probability of 0.05 for running. As a subsequent processing step, the daily activity data acquisition unit records actions with confidence scores exceeding a certain threshold into a time-series database. Further, this unit calculates a baseline of daily rhythm from the recorded action history and calculates a deviation score indicating how much the current action deviates from the baseline. Through this processing, by integrating data from multiple sensors, the system accurately grasps complex daily activities that are difficult to identify with a single sensor. This system effectively quantifies subtle changes in the user's lifestyle patterns, contributing to the early detection of cognitive decline.This system demonstrates its high practicality in specific application areas such as care services for elderly people living alone or self-reliance support systems that utilize smart homes.

[0088] The Communication Data Acquisition Department uses user terminal call history, messaging application usage, SNS usage, voice recognition functions, video calling applications, email applications, camera applications, etc., to obtain data on the number of phone calls with family or friends, the number of messages sent and received, the frequency of SNS posts, the number of video calls, the number of emails sent, the number of photos or videos shared, the length of conversations, the diversity of topics, the presence or absence of emotional expression, the repetition of the same topics, the speed of response to questions, the frequency of topic switching, the presence or absence of jokes or laughter, mispronouncing the other person's name, confusion in family structure, confusion about dates or days of the week, and other aspects of interpersonal communication. For example, voice recognition is used to transcribe daily conversations, and the frequency or time of communication, response speed, tendency to not reply to read messages, and the frequency of stickers or emoticons used are also obtained from SNS or messaging application usage. Furthermore, through video calling application image analysis, changes in facial expressions or eye movements, and emotional expressions such as smiles or surprise are also digitized to evaluate the quality of communication from multiple perspectives. In addition, the frequency or content of conversations with other users or staff, participation in group activities, the intensity of conversations, the number of times someone speaks, and the diversity of their speech are also obtained from the facility's dialogue recording system or recreational activity participation records. Specifically, this communication data acquisition unit has an emotion and intent parsing engine that integrates natural language processing and computer vision technologies. The input data to this engine is unstructured data in multimedia form. Specific examples of input data include speech waveform data acquired from a microphone, arrays of text strings extracted from messaging applications, and tensors of image frames captured from video calls. This engine performs preprocessing on this input data, including textification based on speech recognition, lexical analysis, and facial landmark detection. The output data from this engine is a vector of quantitative communication indicators. Specific examples of output data include emotion score vectors (e.g., joy 0.8, sadness 0.1), numerical values ​​for topic diversity indicators representing the richness of vocabulary used, and the number of times the same phrase appears in a conversation. As a subsequent processing step, this communication data acquisition unit normalizes these indicator vectors and combines them with other datasets as input features for a cognitive function assessment model. This system extracts the quality of communication, which relies on subjective human observation, as objective and continuous numerical data through this processing. It leverages its technical capabilities to precisely capture signs of cognitive decline from subtle changes in the rhythm of conversation or facial expressions. In specific applications such as assistive tools in online medical consultations or communication support robots in nursing facilities, this system makes a significant contribution to understanding the user's mental state.

[0089] The Nursing Record Acquisition Department acquires records of meals, bathing, excretion, movement, falls, forgetfulness, conversations, participation in recreational activities, medication management, vital sign measurements, rehabilitation implementation, accidents, or abnormal behavior entered by nursing staff using tablet terminals or the nursing record system.

[0090] For example, it can record in detail the amount or content of meals, whether or not bathing, the time of bathing, the frequency or condition of excretion, the use of walking aids during movement, the date, time and condition of falls, the frequency or content of forgetfulness, the content or response of conversations, participation in or response to recreational activities, medication administration or forgetting to take medication, vital sign measurements, rehabilitation implementation or achievement, details of accidents or abnormal behavior, etc. Furthermore, nursing staff can also input users' facial expressions or attitudes, abnormalities during meals or excretions, skin condition during bathing, actions or intentions during rehabilitation, nighttime sleep patterns, daytime activity levels, and family meetings in a free-writing manner. This free-writing data also becomes the object of natural language parsing based on generative AI. Specifically, this nursing record acquisition department has a natural language understanding module that converts free-writing text into structured data. The input data to this module consists of text strings entered by nursing staff and selected category data. Specific examples of input data include free-form descriptive text such as "I had no appetite today and felt a bit dazed," and numerical data such as "my food intake was 50%." This module performs named entity extraction and dependency parsing on the input text. The output data from this module is structured event labels and their importance scores. Specific examples of output data include standardized event label strings such as "decreased appetite" or "decreased energy," and importance scores ranging from 0.0 to 1.0, indicating the severity of the event. As subsequent processing, the nursing record acquisition unit appends the generated event labels and importance scores to the corresponding user's record in the integrated database. Furthermore, if the importance score exceeds a predetermined threshold, the nursing record acquisition unit immediately triggers the inference process of the risk prediction model. Through this processing, the system transforms subjective free-form descriptions containing inconsistencies or ambiguities into objective data points that can be processed by the machine learning model. This system effectively extracts important signs buried in nursing records without omission, improving the comprehensiveness and accuracy of data analysis. In specific application areas such as improving the efficiency of nursing record-keeping or automating the creation of nursing plans, this system aims to both reduce the burden on staff and improve the quality of care.

[0091] The Data Integration Department integrates health data, daily activity data, communication data, and nursing record data for each user, generating an integrated dataset managed by time series. For example, it summarizes a single record for a given day, including vital signs, outing history, conversation content, social activity records, medication status, meal content, sleep duration, exercise volume, SNS usage, communication with family, implementation of hobbies, and nursing records, allowing for the tracking of long-term changes or trends. The Data Integration Department also performs data correction for missing or outlier values, anonymization for privacy protection, and data normalization or standardization. For example, it deduplicates step count or sleep data obtained from multiple devices or applications, automatically detects and corrects outliers, and converts all data into a common time series format for management. Specifically, this Data Integration Department possesses an advanced data alignment engine for constructing multidimensional time series tensors. The input data to this engine consists of time series data streams with different sampling rates, asynchronously sent from various acquisition departments. Specific examples of input data include heart rate data recorded at 1-minute intervals, text data from irregularly recorded nursing records, and weight data recorded once a day. This engine performs timestamp synchronization processing on the input data, applying missing value imputation algorithms such as spline interpolation or K-nearest neighbors. The output data from this engine is a multidimensional tensor resampled at regular time intervals. A specific example of the output data is a two-dimensional floating-point tensor with the number of time steps and the number of feature dimensions as axes. The data integration unit, as a subsequent processing step, expands the generated multidimensional tensor into a cached area in memory. This tensor is then provided as input for batch processing of cognitive function assessment models and risk prediction models. Through this processing, the system eliminates the temporal bias of heterogeneous data, making all variables comparable on the same time axis. This system significantly improves the learning efficiency and inference accuracy of machine learning models, while reducing false detections. In specific applicable fields such as building large-scale medical data warehouses or real-time health status monitoring platforms, this system plays a core role as the data foundation.

[0092] The Cognitive and Physical Function Assessment Department analyzes and integrates datasets, utilizing generative AI to calculate cognitive and physical function scores from users' daily conversations, behavior patterns, health data, and care records. For example, it uses generative AI-based natural language processing technology to analyze conversation content in detail, incorporating indicators such as repetition of the same topic or pauses in speech, response speed to questions, vocabulary diversity, logical coherence, richness of emotional expression, use of jokes or metaphors, accurate recognition of the other person's name or date, consistency of conversation, frequency of topic switching, topic deviation, and expressions of care for the other person. Furthermore, it reflects changes in behavior patterns (a sharp decrease in the frequency of going out, a decline in willingness to participate in activities of interest, increased forgetfulness, disrupted sleep rhythms, reduced meal frequency, increased forgetfulness of medication, and errors in housework steps) in the scoring. Additionally, it calculates physical function scores based on walking speed or range of movement, number of falls, progress of rehabilitation, food intake, bowel movements, bathing frequency, amount of exercise, and stability of vital signs. For example, a decrease in walking speed lowers the physical score, and an increase in the number of falls increases the risk. The generative AI is provided with prompts such as "Extract signs of changes or declines in cognitive or physical function from the user's conversation history, activity records, health data, and care records over the past month, and score them," and "Detect patterns such as repetitive conversations or delayed responses, reduced vocabulary, decreased walking speed, increased number of falls, and reduced food intake, and assess the risk." Specifically, this assessment department has an advanced score calculation engine using multimodal neural networks. The input data to this engine is a multidimensional tensor of an integrated dataset supplied by the data integration department. Specific examples of input data include matrix data representing behavioral features over the past 30 days, or text embedding vectors of conversations. This engine feeds this input data into multiple specialized sub-networks, extracts the hidden state vectors of each modality, and combines them. The output data from this engine are cognitive and physical function scores as continuous values, along with confidence intervals for each score. Specific examples of output data include floating-point numbers such as a cognitive function score of 75.5 and a physical function score of 68.2, along with confidence intervals of ±3.1. As a follow-up process, this assessment department compares the calculated scores with past performance data to calculate the rate of change. If the score decreases by more than 10% compared to the previous week, an alarm is set to indicate a sharp decline in function. Through this processing, by integrating multiple modalities for assessment, this system can accurately detect subtle functional declines that might be missed in a single test. This system provides a continuous stream of objective indicators to supplement the diagnosis of specialists. It offers significant clinical value in specific application areas such as early screening tools for dementia and systems for measuring the effectiveness of rehabilitation.

[0093] The Risk Prediction Department uses a generative AI-based machine learning model to comprehensively analyze family or past medical history, changes in cognitive and physical scores, fall history, medication status, frequency of social interaction, and changes in eating or sleeping patterns to individually predict the risk of dementia or physical function decline. For example, the generative AI is provided with prompts such as "Predict the risk of dementia or physical function decline based on this user's integrated dataset and explain the risk factors" and "Extract lifestyle habits or behavior patterns that contribute to the increased risk and create improvement proposals," determining the risk level (low, medium, high) and outputting risk factors or improvement points in natural language. For instance, in the case of a sharp decline in cognitive function scores and an increase in the number of falls, a "high risk of dementia and fall" is determined, automatically generating improvement proposals such as "strengthen accompaniment when going out," "implement walking aids," and "recommend a specialist consultation." Specifically, this Risk Prediction Department has an ensemble learning engine that integrates gradient boosting decision trees or prediction models based on self-attention mechanisms. The input data to this engine is a feature vector containing the user's static and dynamic features. Specific examples of input data include one-hot vectors representing family or past medical history, and time-series vectors representing the shift in cognitive scores over the past 6 months. This engine performs feature scaling on these input data before feeding them into the predictive model for inference. The output data from this engine consists of categorical values ​​for the probability of future functional decline, risk levels, and scores representing the importance of each feature. Specific examples of output data include a floating-point number indicating a fall probability of 0.78 after 6 months, a string indicating a high risk level, and a numerical vector indicating the importance of sleep deprivation of 0.45. As a subsequent processing step, the risk prediction department sends a control signal to the care plan generation department instructing them to prioritize the creation of an intervention plan when the output risk level is high. Simultaneously, the risk prediction department generates natural language text explaining the risk factors based on the feature importance. Through this processing, the system imbues the predictions of AI, which can easily become black boxes, with mathematically based interpretability. This system aims to increase the confidence of caregivers and families in the prediction results, promoting a transition to appropriate preventative actions. This system provides strong support for data-driven decision-making in specific application areas such as risk stratification tools in preventive medicine or payment optimization systems for long-term care insurance.

[0094] The nursing plan generation department, based on risk prediction results or users' interests (such as music, sports, gardening, cooking, reading, crafts, Go, Shogi, karaoke, pets, travel, etc.), living conditions, family structure, and past activity history, provides generative AI with prompts such as "creating a one-week rehabilitation menu, life support suggestions, entertainment activity proposals, and feedback for family and staff based on the user's risk and interests," and "generating memory training content using the user's photos or memories," automatically generating individually optimized nursing plans. For example, in high-risk situations, it strongly recommends medical institution visits or specialist consultations; suggests music therapy or singing activities for users who enjoy music; and suggests individualized rehabilitation or walking programs for users who enjoy sports. Furthermore, it automatically generates information such as "recent changes in cognitive or physical function" and "key points to note in the future" for family members or caregivers. Feedback includes examples of conversations in daily life. For instance, suggestions are generated for families such as, "Recently, repetitive conversations have been increasing, so it's best to start conversations in a slow tone and provide specific opportunities when changing the topic." Specifically, this nursing plan generation department has a scheduling engine that uses reinforcement learning algorithms to optimize the plan. The input data to this engine includes risk prediction results, user profile vectors, and historical data on responses to past plans. Specific examples of input data include a fall risk score, a multidimensional vector encoding interests, and a percentage representing past participation in music therapy. This engine defines this input data as a state space and performs optimization calculations by setting the expected improvement as the reward function. The output data from this engine is structured data defining a one-week schedule and expected performance scores for each task. Specific examples of output data include a JSON-formatted task list array containing the start time, end time, and content of each task, and a numerical value representing the expected improvement in cognitive function based on that task. As a subsequent processing step, this nursing plan generation department sends the generated schedule data to the nursing staff's terminal. This nursing plan generation department accepts feedback or corrections from staff, using the results as rewards for the reinforcement learning model to update its weights. Through this process, the system continuously learns from users' preferences and past response history, improving the accuracy of personalization. This system effectively generates highly consistent nursing plans, significantly reducing the time nursing managers need to create plans. In specific application areas such as individual function training and calculation support systems or automated recreational activity proposal systems in day care services, this system simultaneously improves operational efficiency and service quality.

[0095] The AI-powered Nursing Counseling Department uses tablets or smartphones, voice assistants, chatbots, and other devices to propose nursing plans, rehabilitation menus, life support suggestions, and feedback to users or caregivers, monitoring implementation and user reactions. For example, it offers specific action proposals or raises awareness such as "Let's do 10 minutes of music therapy today," "Let's take a walk and stretch this afternoon," "Recent conversations have decreased, so try video calling your family," and "Falls are increasing, so explore the use of walking aids." When users or staff input implementation details or reactions, this information is again accumulated in the data integration department and used in future assessments or plan generation.

[0096] Furthermore, generative AI is used to analyze users' activity implementation, emotional expression, lifestyle rhythm, changes in physical condition, and feedback from family or caregivers in real time. This dynamically generates and provides encouragement, reminders, attention-grabbing suggestions, and action proposals at the optimal time and method for each user, automatically updating the care plan or rehabilitation menu based on the user's response or implementation. Specifically, this AI-powered nursing guidance department has an intervention management module that executes context-aware notification control algorithms. Input data to this module includes the current time, the user's current location, current activity status, and task data from the care plan. Specific examples of input data include a timestamp indicating the current time, GPS coordinates, a status flag indicating whether the user is asleep or awake, and the scheduled task ID. Based on this input data, the module calculates the time when the user is most likely to accept the notification. Output data from this module includes the notification execution time, the optimal notification channel, and the generated message text. Specific examples of output data include a timestamp indicating execution in 5 minutes, a channel identifier specifying the speaker, and a string of natural language prompting a walk. As subsequent processing, this AI-powered nursing guidance department calls the target device's API at the specified time to execute the notification. Simultaneously, this AI-powered nursing guidance system records user responses to notifications and measures response time. This process eliminates the annoyance of indiscriminate notifications that ignore user circumstances. The system effectively promotes action change by intervening at the optimal time and in the most appropriate manner to increase proposal acceptance. This system achieves sustained user engagement in specific application areas such as applications for improving medication adherence or action support systems for preventing lifestyle-related diseases.

[0097] In addition, regular reassessments of cognitive or physical function are conducted, and generative AI is used to analyze changes in scores or risk levels, providing prompts such as "summarizing the user's activity history and cognitive and physical function score trends over the past month, creating a report for the family" and "summarizing risk factors and recommending intervention strategies for medical institutions," automatically generating optimal feedback or reports for families, medical institutions, and caregivers. In cases of a sharp decline in cognitive or physical function, or in the event of a fall or accident, the generative AI automatically summarizes the content and sends alerts such as "recommendation to a specialist" and "please strengthen fall prevention procedures." For example, if the number of nighttime urination increases sharply while sleep time decreases, it is determined that "the risk of falls is increased due to nighttime urination frequency," automatically proposing measures such as enhancing nighttime lighting, ensuring proper toilet access routes, recording urination diaries, and recommending urological visits. Specifically, this system has a report generation engine that uses natural language generation models to automate document creation. The input data to this engine includes score trends over a certain period, anomaly detection logs, and historical data of implemented care plans. Specific examples of input data include an array of cognitive function scores over the past month, alarm history from fall detection sensors, and a list of completed rehabilitation tasks. This engine maps this input data to templates, generating natural sentences appropriate to the context. The output data from this engine consists of structured report data and urgency indicators. Specific examples of output data include tagged text for PDF generation and Boolean values ​​indicating high urgency. As a subsequent process, the system determines whether the urgency indicator is true. If the indicator is true, the system immediately sends the report to the designated contact method using a secure communication protocol. Through this processing, the system fully automates the process of creating standardized reports. This system effectively creates time for caregivers to focus on direct care and prevents delays in information sharing. In specific applications such as information-sharing platforms for multi-professional collaboration or regular report distribution services for families, this system drastically reduces communication costs.

[0098] Furthermore, an external collaboration department is established to collaborate securely with medical institutions, care providers, regional integrated support centers, families, social welfare workers, care managers, pharmacists, nutritionists, rehabilitation personnel, and regional event organizers to share necessary information. For example, with the user's consent, cognitive function scores or risk levels and activity records can be sent to medical institutions; care providers can collaborate to adjust care plans; regional activity information can be automatically retrieved and proposed to the user; and family members or medical institutions can be automatically notified in emergencies such as falls or sudden cognitive decline. For example, information on regional health classrooms or interest groups, volunteer activities, rehabilitation classrooms, dementia cafes, etc., can be automatically collected, and the best activities can be proposed based on the user's interests or risk. Specifically, this external collaboration department has a secure API gateway and a data encryption processing module. Input data to this module includes datasets of shared objects, access control policies, and API specification data from external systems. Specific examples of input data include records containing the user's cognitive function score, policy rules allowing access to the medical institution, and the endpoint URL of the destination. This module applies advanced encryption algorithms to this input data to generate the data payload. The output data from this module is an encrypted JSON payload and an authentication token. Specific examples of the output data include encrypted string data and an expired JWT token. As a subsequent process, this external collaboration unit sends POST requests to the RESTful API endpoints of external systems via HTTPS. This external collaboration unit verifies the received response code and logs the success or failure of the transmission. Through this process, the system ensures interoperability between systems while completely preventing the leakage of highly confidential medical and nursing data. This system enables robust data collaboration, allowing for the construction of a comprehensive nursing network across the region. In specific applicable areas such as data collaboration with electronic health record systems or integration with autonomous region health improvement platforms, this system functions as a secure data hub.

[0099] Thus, the embodiments of the present invention possess a multifaceted and robust structure, including collaboration between servers and user terminals, multi-layered and dynamic utilization of generative AI, optimized intervention and continuous support for each user, real-time data analysis and feedback, collaboration with medical, nursing, family, and regional resources, emergency response, privacy protection, and improved user experience. This comprehensively achieves improved quality of care, support for user self-reliance, reduced burden on families or staff, controlled medical costs, and extended healthy lifespan. Specifically, this embodiment features an orchestration engine that comprehensively manages all data streams processed throughout the system. Input data to this engine consists of system usage metrics and resource usage data collected from each subsystem. Specific examples of input data include arrays of values ​​in milliseconds representing the inference latency of each AI model, or percentages representing the CPU utilization of the server. This engine monitors this input data in real time and executes load balancing algorithms. Output data from this engine consists of optimized system usage parameters and scaling instructions. Specific examples of output data include values ​​representing the traffic allocation rate to a specific processing node, and control commands instructing the startup of new virtual machines. This system, as a subsequent processing step, calls upon the cloud infrastructure's APIs to perform automatic scaling control. Through this processing, the system maintains a scalable distributed processing architecture, responding to sudden increases in data volume without delay. This system effectively supports a large number of users simultaneously, contributing to the resolution of societal nursing challenges. In specific applicable areas such as national-level healthcare data infrastructure or globally deployed digital health services, this system provides extremely high availability and performance.

[0100] (System Composition) The system involved in this embodiment includes a health data acquisition department, a daily activity data acquisition department, a communication data acquisition department, a nursing record acquisition department, a data integration department, a cognitive and physical function assessment department, a risk prediction department, a nursing plan generation department, an AI nursing guidance department, and an external collaboration department.

[0101] The Health Data Acquisition Department automatically acquires various health-related data from wearable devices such as smartwatches, activity meters, blood pressure monitors, thermometers, pulse oximeters, and blood glucose meters, as well as from health management applications, electronic medical records from regular physical examinations, and medication management applications installed on smartphones or tablets. This data includes heart rate, blood pressure, body temperature, steps, sleep duration, weight, blood glucose levels, oxygen saturation, calories burned, resting heart rate, exercise intensity, wake-up and bedtime times, past medical history, family medical history, medication history, allergy information, and results of regular physical examinations. For example, it can acquire daily morning blood pressure measurements from smartwatches via Bluetooth, weekly step counts from activity meters, monthly weight changes from scales via Wi-Fi, and medication status from medication management applications. It can also automatically link to electronic medical records from medical institutions, including blood test results, imaging diagnostics, and doctor's opinions. Furthermore, it integrates data from multiple medical institutions and automatically detects and corrects duplicate or contradictory data. Specifically, this health data acquisition unit includes a preprocessing pipeline that uses algorithms to complete data gaps caused by sensor noise or communication errors. The data input to this pipeline is raw, unprocessed data streams sent from various devices. Examples of input data include time-series arrays of heart rates sampled at 1 Hz, or numerical data of blood glucose values ​​recorded at irregular intervals. This pipeline applies moving average filters or outlier removal algorithms to this input data. The output data from this pipeline is a clean data vector that has undergone noise removal and interpolation. Examples of output data include continuous time-series vectors with missing values ​​linearly interpolated, or numerical arrays with standardized units. This clean data vector is then passed to a feature extraction module for further processing. Through this processing, the system eliminates uncertainties in the data caused by hardware limitations in the software. This system enhances the robustness of subsequent AI processing and prevents erroneous inferences based on incorrect data. This system ensures data reliability in specific application areas such as open platforms integrating wearable devices from multiple manufacturers, or remote monitoring systems for home-based patients.

[0102] The Daily Activity Data Acquisition Department uses the user terminal's GPS function or accelerometer, gyroscope, calendar applications, activity tracking applications, voice input applications, image recognition applications, etc., to acquire data on daily activity patterns such as frequency of outings, distance traveled, amount of exercise, number of times shopping or participating in hobbies, completion of household chores, sleep patterns, meal times, bathing times, number of toilet visits, TV viewing time, reading time, whether or not napping, diversity of walking routes, and use of transportation. For example, it integrates with calendar applications to automatically record the input, execution, and forgetting of scheduled activities; automatically determines outing destinations or travel routes from GPS data; analyzes exercise intensity or walking patterns from accelerometer data; and uses image recognition applications to manage food in the refrigerator, identify medication packaging, and confirm garbage disposal status. Furthermore, it also collects user input in diary applications or activity records entered via voice input, such as subjective activity records like "I took a walk in the nearby park today," "I watched TV for 2 hours after lunch," and "I prepared dinner myself," which are also accumulated as data. In addition, data such as the number of times a room is entered and exited, the number of times someone goes to the toilet at night, the number of times someone gets out of bed, and changes in room temperature or humidity are also collected from IoT devices such as door open / close sensors, bed sensors, lighting sensors, and temperature and humidity sensors within the facility to monitor daily routines or signs of abnormal behavior. Specifically, this daily activity data acquisition unit has an action segmentation engine that can formulate complex daily activities into discrete state transition models. The data input to this engine consists of spatial coordinate data, acceleration tensors, and numerical arrays from environmental sensors. Specific examples of input data include matrix data representing latitude and longitude trajectories, tensor data representing 3-axis acceleration changes, and floating-point arrays representing room temperature changes. This engine applies Hidden Markov Models or Recurrent Neural Networks to this input data to segment continuous data into meaningful action units. The output data from this engine consists of pairs of start and end times for each action segment, along with the inferred action category. Specific examples of output data include timestamps for the start and end times, and labels for action categories such as eating or sleeping. This daily activity data acquisition unit stores these action segments in a graph database for further processing. This daily activity data acquisition unit calculates a probability matrix of behavioral transitions from stored data, monitoring for any unusual behavioral transitions. Through this processing, the system extracts human intent or life context from a simple list of sensor values. This system enables the mathematical detection of abnormal behavioral patterns, playing a crucial role in the early detection of cognitive decline. In specific applications such as control systems for life support robots for the elderly and resident care infrastructure in smart cities, this system provides a high level of situational understanding.

[0103] The Communication Data Acquisition Department uses user terminal data such as call history, messaging application usage, SNS usage, voice recognition functions, video calling applications, email applications, and camera applications to collect data on the number of phone calls, messages sent and received, SNS posting frequency, video calls, emails sent, photo or video sharing frequency, conversation length, topic diversity, emotional expression, repetition of the same topic, response speed to questions, topic switching frequency, presence of jokes or laughter, mispronouncing names, family composition inconsistencies, and date or day-of-week inconsistencies to assess the content or frequency of interpersonal communication. For example, it uses voice recognition to transcribe daily conversations and also obtains data on communication frequency or time periods, response speed, read-but-not-reply tendencies, and the frequency of sticker or emoticon usage from SNS or messaging application usage. Furthermore, it uses video analysis from video calling applications to digitize changes in facial expressions or eye movements, and emotional expressions such as smiles or surprise, to evaluate the quality of communication from multiple perspectives. In addition, the frequency or content of conversations with other users or staff, participation in group activities, the intensity of conversations, the number of times someone speaks, and the diversity of their speech are also obtained from the facility's dialogue recording system or entertainment activity participation records. Specifically, this communication data acquisition unit employs a multi-threaded architecture that processes multiple input streams in parallel. This acquisition unit performs a Fast Fourier Transform on the speech data. This acquisition unit converts the time-domain speech signal into a frequency-domain spectrogram tensor. This acquisition unit inputs this spectrogram tensor into a convolutional neural network. This network extracts speech features. This acquisition unit inputs the extracted features into a recurrent neural network. This network parses the context of the time series. This acquisition unit generates spoken text data from the parsing results. This acquisition unit inputs the generated text data into a large language model. The data input into this model is word embedding vectors. The dimension of the data input into this model is, for example, 768 dimensions. Specific examples of the data input into this model include text vectors representing the content of user speech or timestamp arrays representing the timing of speech. This model applies a self-attention mechanism to the input vector group. This model calculates the dependencies between words. This model outputs an emotion polarity score or a topic diversity score. The output data from this model is in the form of an array of floating-point numbers. Specific examples of the output data from this model include a score of 0.85 representing joy and a score of 0.12 representing sadness. This acquisition unit compares these output scores with a predetermined threshold. If the acquisition unit detects an emotion change exceeding the threshold, it sends a flag indicating that change to the subsequent data integration unit. This acquisition unit performs the same processing on image data. The acquisition unit extracts image tensors frame by frame from the animated images captured by the camera. This image tensor is input into the facial organ detection model. This model outputs a feature point coordinate vector representing the movement of the eyes or mouth.This acquisition unit quantifies the degree of facial expression changes based on the coordinate vector. This acquisition unit stores these numerical data in a time-series database. Through this structure, the system automatically extracts objective communication indicators independent of human subjectivity. This system eliminates assessment bias. This system achieves the technical effect of highly accurate capture of subtle changes in cognitive function. This technology is extremely useful in specific application areas such as daily care systems in elderly facilities or consultation support systems in telemedicine.

[0104] The Nursing Record Acquisition Department is responsible for acquiring records entered by nursing staff using tablet terminals or a nursing record system, including information on meals, bathing, excretion, movement, falls, forgetfulness, conversations, participation in recreational activities, medication administration, vital sign measurements, rehabilitation implementation, and incidents or abnormal behaviors. For example, it can record detailed information such as food intake or content, the presence or timing of bathing, the frequency or condition of excretions, use of walking aids during movement, the date, time, and condition of falls, the frequency or content of forgetfulness, the content or response to conversations, participation in recreational activities or responses, medication administration or forgetting to take medication, vital sign measurements, rehabilitation implementation or achievement, and detailed information on incidents or abnormal behaviors. Furthermore, nursing staff can also freely input user facial expressions or attitudes, abnormalities during meals or excretions, skin condition during bathing, actions or intentions during rehabilitation, nighttime sleep patterns, daytime activity levels, and family meetings. This freely recorded data also becomes the object of natural language processing based on generative AI. Specifically, this nursing record acquisition department possesses a natural language processing pipeline that converts free-form text, as unstructured data, into structured data. The acquisition department inputs the text data into a morphology parsing engine. This engine segments the text into word tags. The acquisition department performs preprocessing to remove unwanted stop words. The acquisition department inputs the preprocessed tag list into a large language model. The data input to this model is an index array of tokenized text. The maximum dimension of the data input to this model is 512 tags. Specific examples of data input to this model include text tags representing choking during meals or text tags representing wandering at night. This model features a multi-layer encoder based on a Transformer architecture. This model generates a high-dimensional feature vector that considers context from the input tag list. This feature vector is input into a fully connected layer. This model classifies the content of the nursing record into multiple predefined categories. The output data from this model is in the form of a Softmax distribution representing the probability of belonging to each category. As a specific example of the output data from this model, there is a probability value of 0.92 for the fall risk category and a probability value of 0.78 for the cognitive decline category. This acquisition unit assigns the category with the highest probability as the label for this record. This acquisition unit associates the extracted label with the original text data and stores it in a relational database. This acquisition unit transmits this label information to the subsequent risk prediction unit. Through this structure, this system extracts important signs without omission from a large amount of free-form descriptive data. This system achieves a significant improvement in the accuracy of data analysis while reducing the workload of nursing staff in record-keeping. This technology is extremely useful in specific application areas such as business support systems in large nursing facilities or family-oriented information sharing platforms in home care.

[0105] The Data Integration Department integrates health data, daily activity data, communication data, and nursing record data for each user, generating integrated datasets managed by time series. For example, it aggregates daily vital sign information, outing history, conversation content, social activity records, medication status, meal content, sleep duration, exercise volume, SNS usage, communication with family, implementation of hobbies, and nursing records into a single record, enabling the tracking of long-term changes or trends. The Data Integration Department also performs data correction for missing or outlier values, anonymization for privacy protection, and data normalization or standardization. For example, it deduplicates step count or sleep data obtained from multiple devices or applications, automatically detects and corrects outliers, and converts all data into a common time series format for management. Specifically, this Data Integration Department possesses a multimodal integration engine that maps heterogeneous, mixed data to a common latent space. This integration department receives data streams asynchronously sent from various acquisition departments. This integration department performs timestamp-based synchronization processing on the received data. This integration department performs autoencoder-based completion processing on time series data containing missing values. This integration unit inputs incomplete data into an autoencoder that has learned normal data patterns. This integration unit estimates the missing values ​​in a way that minimizes reconstruction error. This integration unit executes a prediction model using a Long Short-Term Memory network for outlier detection. This integration unit predicts current values ​​from past time-series data. This integration unit calculates the deviation between the predicted and actual observed values. This integration unit excludes or smooths data with deviations exceeding a specified threshold as outliers. This integration unit converts the preprocessed data into a multidimensional time-series matrix. The data input to this integration unit is an array of numerical vectors or categorical variables with different sampling rates. The dimensionality of the data input to this integration unit can reach tens to hundreds of dimensions. Specific examples of the data input to this integration unit include heart rate vectors at one-minute intervals or weight data collected once a day. This integration unit resamples this data to a common temporal resolution. The output data from this integration unit is in the form of a normalized multidimensional tensor. Specific examples of the output data from this integration unit include a 128-dimensional feature tensor summarized over a day, or an activity matrix over an hour. This integration department stores the generated integrated dataset in a distributed file system. Responding to queries from subsequent evaluation departments, this integration department provides the dataset at high speed. Through this structure, the system automatically generates high-quality analytical data from real-world environmental data that contains noise or missing information. This system significantly improves the inference accuracy of subsequent artificial intelligence models. This technology is extremely useful in specific application areas such as integrated management systems for personal health records or as a foundation for medical data sharing among multiple hospitals.

[0106] The Cognitive and Physical Function Assessment Department analyzes and integrates datasets, utilizing generative AI to calculate cognitive and physical function scores from users' daily conversations, behavior patterns, health data, and care records. For example, it uses generative AI-based natural language processing technology to analyze conversation content in detail, considering indicators such as repetition of the same topic or pauses in speech, response speed to questions, vocabulary diversity, logical connections, richness of emotional expression, use of jokes or metaphors, accurate recognition of the other person's name or date, consistency of conversation, frequency of topic switching, topic deviation, and expressions of care for the other person. Furthermore, it reflects changes in behavior patterns (a sharp decrease in the frequency of going out, a decline in willingness to participate in activities of interest, increased forgetfulness, disrupted sleep rhythms, reduced meal frequency, increased forgetfulness of medication, and errors in housework steps) in the scoring. Physical function scores are calculated based on walking speed or range of movement, number of falls, progress of rehabilitation, food intake, bowel movements, bathing frequency, amount of exercise, and stability of vital signs. The generative AI is provided with prompts such as "Extract signs of changes or declines in cognitive or physical function from the user's conversation history, activity records, health data, and care records over the past month, and score them," and "Detect patterns such as repetitive dialogue or delayed response, reduced vocabulary, decreased walking speed, increased number of falls, and reduced food intake, and assess the risks" for analysis. Specifically, this cognitive and physical function assessment department has an integrated inference architecture that combines a large-scale language model and a time-series analysis model. This assessment department obtains an integrated dataset from the data integration department. This assessment department inputs this dataset into the prompt generation module. This module converts numerical or categorical data into natural language text expressions. This module combines the converted text with the prescribed instructions to generate prompt strings. This assessment department inputs the generated prompt strings into a large-scale language model. The data input into this model is a tokenized array of prompt indices. The maximum dimension of the data input into this model is 4,096 tokens. As specific examples of the data input into this model, there are text tokens indicating the progression of steps over the past week, or text tokens indicating a decrease in vocabulary in the conversation. This model is a Transformer network with billions of parameters. It uses pre-learned language understanding capabilities to deeply analyze the context of the input text. The model infers scores representing the states of cognitive and physical function. The output data from this model is structured JSON text data. Specific examples of the output data include a value representing a cognitive function score of 85 or a physical function score of 72. The evaluation department performs syntactic parsing on the output text data. The evaluation department extracts the numerical scores and records them in a database. The evaluation department then sends the extracted scores to the subsequent risk prediction department.Through this architecture, the system achieves a level of understanding of complex contexts that was previously difficult to attain with purely rule-based assessments. This system automates high-precision functional assessments, bringing them closer to those of specialist diagnoses. This technology is extremely useful in specific application areas such as screening systems in telemedicine or healthy life extension programs for the elderly in autonomous communities.

[0107] The Risk Prediction Department uses a generative AI-based machine learning model to comprehensively analyze family or past medical history, changes in cognitive and physical scores, fall history, medication status, frequency of social interaction, and changes in eating or sleeping patterns to individually predict the risk of dementia or physical function decline. For example, it provides the generative AI with prompts such as "Predict the risk of dementia or physical function decline based on the user's integrated dataset and explain the risk factors" and "Extract lifestyle habits or behavior patterns that contribute to the increased risk and create improvement proposals," determining the risk level (low, medium, high) and outputting risk factors or improvement points in natural language. For instance, in the case of a sharp decline in cognitive function scores and an increase in the number of falls, it is determined to be "high risk of dementia and fall," automatically generating improvement proposals such as "strengthen accompaniment when going out," "implement walking aids," and "recommend a specialist consultation." Specifically, this Risk Prediction Department has a hybrid prediction engine that combines a gradient boosting decision tree algorithm and a deep neural network. This prediction department obtains various scores calculated by the assessment department and feature quantities extracted from the integrated dataset. This prediction unit combines this data to generate a high-dimensional feature vector. This feature vector is then input into a deep neural network. The data input to this network is a one-dimensional array of normalized floating-point numbers. The data input to this network has 256 dimensions. Specific examples of the input data include a moving average of cognitive function scores over the past month, or binary flags representing fall history. This network learns the non-linear relationships of the input data through multiple hidden layers. This network uses Dropout layers to suppress overfitting. This network uses the Softmax function of the final layer to calculate the probability of each risk level. The output data from this network is in the form of a probability distribution vector for low, medium, and high risk classes. Specific examples of the output data from this network include a probability value of 0.82 for high risk and a probability value of 0.15 for medium risk. This prediction unit determines the class with the highest probability as the final risk level. This prediction unit uses a gradient boosting decision tree model to calculate the contribution of each feature to the risk. This prediction unit identifies features with high contributions as risk factors. This prediction unit inputs the identified risk factors into a large-scale language model. This model generates specific improvement proposal texts based on risk factors. The prediction department then sends the determined risk level and the generated improvement proposals to the subsequent care plan generation department. Through this structure, the system clarifies the judgment criteria of artificial intelligence, which is easily subject to manipulation. This system achieves high-precision risk prediction while ensuring transparency and reliability in risk forecasting. This technology is extremely useful in specific application areas such as the selection system for preventative benefits in long-term care insurance, or the assessment system for health enhancement insurance from private insurance companies.

[0108] The nursing plan generation department, based on risk prediction results or the user's interests, lifestyle, family structure, and past activity history, provides generative AI with prompts such as "create a one-week rehabilitation menu, life support suggestions, entertainment activity proposals, and feedback for family and staff, tailored to the user's risk and interests," and "generate memory training content using the user's photos or memories," automatically generating individually optimized nursing plans. For example, in high-risk situations, it strongly recommends medical institution visits or specialist consultations; suggests music therapy or singing activities for users who enjoy music; and suggests individualized rehabilitation or walking programs for users who enjoy sports. Furthermore, it automatically generates feedback for family and caregivers, such as "recent changes in cognitive or physical function," "key points to pay attention to in the future," and "examples of conversation in daily life." Specifically, this nursing plan generation department has a planning engine that uses a generative model optimized through reinforcement learning. This generation department obtains risk levels and improvement proposals from the risk prediction department. This generation department obtains data on the user's interests or family structure from the user profile database. This generation department integrates this information to generate a context vector. This generation department inputs this context vector into a generator of a conditional generative adversarial network. The data input to this generator is a combined tensor of context vectors and random noise vectors. The data input to this generator has 512 dimensions. Specific examples of the input data include embedding vectors representing categorical variables such as music preference or numerical vectors representing high-risk states. This generator generates multiple candidate weekly care plans based on the input data. The generation department inputs the generated plan candidates into a discriminator. The discriminator evaluates the effectiveness of each plan based on past success data. The generation department selects the plan with the highest evaluation score from the discriminator as the optimal plan. The generation department inputs the selected plan into a large language model. This model converts the plan content into natural language text or structured data. The output data from this model is in the form of a hierarchical data structure containing date and time, activity content, and precautions. Specific examples of the output data from this model include schedule data for music therapy at 10:00 AM on Mondays or text data about preventing falls while walking. The generation department sends the generated care plan to the subsequent AI-powered nursing guidance department. This system, through its architecture, does not dynamically generate generic care plans, but rather dynamically generates customized plans that are fully adapted to individual characteristics or changes in circumstances. This system delivers a significant technological boost to the quality of care services. This technology is extremely useful in specific application areas such as care manager support systems for creating care plans, or automated scheduling applications for home care.

[0109] The AI-powered Nursing Counseling Department uses tablets or smartphones, voice assistants, chatbots, etc., to suggest nursing plans or rehabilitation menus, life support suggestions, and provide feedback to users or caregivers, while monitoring implementation and user responses. For example, it offers specific action proposals or attention-grabbing suggestions such as "Let's do 10 minutes of music therapy today," "Let's take a walk and stretch this afternoon," "We've been talking less lately, let's try video calling with family," and "Falls have increased, let's explore using walking aids." When users or staff input implementation details or responses, this information is again accumulated in the data integration department and used for future assessments or plan generation. Furthermore, using generative AI, it analyzes user activity implementation, emotional expression, lifestyle rhythm, changes in physical condition, and feedback from family or caregivers in real time, dynamically generating and providing encouragement, reminders, attention-grabbing suggestions, and action proposals at the optimal time or method for each user, and automatically updating nursing plans or rehabilitation menus based on user responses or implementation. Specifically, this AI-powered Nursing Counseling Department has a dynamic intervention engine that integrates real-time stream processing and reinforcement learning agents. The department receives optimized nursing plans from the nursing plan generation department. This guidance unit continuously acquires real-time biometric or behavioral data from the user's terminal or environmental sensors. This unit converts the acquired real-time data into a state vector. This state vector is then input into a deep reinforcement learning model. The data input to this model is a multi-dimensional numerical vector representing the current user state. The data input to this model has 64 dimensions. Specific examples of the input data include a value representing the current heart rate or a value representing the activity level of the past hour. This model uses a policy gradient method based on a Markov decision process to calculate the optimal action. This model determines the timing and content of the intervention that maximizes the user's motivation and minimizes risk. The output data from this model is in the form of a probability distribution of each action selection in a predefined action space. Specific examples of the output data from this model include a probability of 0.75 for executing a voice encouragement or a probability of 0.20 for executing a text notification displayed on the screen. This unit selects and executes the action with the highest probability. This unit acquires the user's response to the executed intervention from sensor data or voice input. This response is used as a reward signal for calculation. This guidance department updates the weight parameters of the reinforcement learning model in real time using calculated reward signals. When a user's state changes beyond a predetermined threshold, this guidance department sends a trigger signal to the care plan generation department requesting the regeneration of the care plan. Through this structure, the system does not execute static schedules but instead achieves adaptive interventions based on the user's instantaneous state. This system dramatically improves the success rate of interventions and user compliance. This technology is extremely useful in specific application areas such as medication support robots for patients with chronic diseases or personal training systems in gyms.

[0110] The External Collaboration Department collaborates with healthcare institutions, care providers, regional integrated support centers, families, social welfare workers, care managers, pharmacists, nutritionists, rehabilitation workers, and regional event organizers to securely share necessary information. For example, with the user's consent, it sends cognitive function scores or risk levels and activity records to healthcare institutions; collaborates with care providers to adjust care plans; automatically retrieves regional activity information and proposes it to the user; and automatically notifies families or healthcare institutions in emergencies such as falls or sudden cognitive decline. For example, it automatically collects information on regional health classrooms or interest groups, volunteer activities, rehabilitation classrooms, dementia cafes, etc., and proposes the best activities based on the user's interests or risk. Specifically, this External Collaboration Department has an information-sharing gateway that combines secure communication protocols and distributed ledger technology. This collaboration department acquires assessment scores or care plan data generated within the system. This collaboration department encrypts the data to be sent using advanced encryption standards algorithms. This collaboration department generates application programming interface requests for the system to send the encrypted data to external healthcare institutions or care providers. The data input to this collaboration department is a data packet containing an identifier of the destination and an encrypted payload. The data input to this collaborative department is a variable-length byte array. Specific examples of input data include JSON-formatted strings containing a user's latest cognitive function score or emergency alert flags. This collaborative department uses a mutual authentication protocol to verify the legitimacy of the communicating party. Data is only sent upon successful verification. The department records the data transmission history or access logs on a blockchain network. This department prevents data tampering and ensures traceability. This department executes an external web scraping module. This department uses natural language processing to automatically extract the dates, times, or content of events from regional event information websites. The output data from this collaborative department is a structured database record of the extracted event information. Specific examples of output data include location coordinates representing the venue of a dementia café or text data indicating participation conditions. This collaborative department forwards the extracted event information to the Artificial Intelligence Nursing Counseling Department. Through this structure, the system achieves seamless information sharing among multiple professions while strictly protecting sensitive personal information. This system effectively strengthens its function as the digital foundation of a regional integrated nursing system. This technology is extremely useful in specific application areas such as medical information collaboration network systems or welfare service integration platforms for autonomous regions.

[0111] Thus, the system described in this embodiment, through close collaboration among its various components and multi-layered, dynamic utilization of generative AI, comprehensively achieves optimized nursing activities and continuous support for each user, real-time data analysis and feedback, collaboration with medical, nursing, family, and regional resources, emergency response, privacy protection, and improved user experience. Specifically, this system adopts a distributed, collaborative architecture combining high-speed networks with edge computing devices and cloud server clusters. Each processing unit is containerized as an independent microservice. The system uses container orchestration tools to dynamically allocate computing resources based on the load of each service. The system executes lightweight inference models at the edge. The system performs high-real-time processing with zero latency. The system performs highly efficient batch processing using large language models or deep learning models on the cloud. The system enables detailed analysis based on massive amounts of historical data. The data input to this system consists of millions of sensor data streams sent per second from thousands of users. The dimensionality of the data input to this system reaches tens of thousands of dimensions overall. Specific examples of the data input to this system include real-time heart rate matrices for all users or tensor streams of full-camera imagery within the facility. This system processes large datasets at high speed in an in-memory database. Output data from this system takes the form of optimized intervention command sets for individual users and statistical dashboards for administrators. Specific examples of output data include emergency notification commands to specific users or fall risk distribution maps for the entire facility. Through these outputs, the system provides comprehensive care across the boundaries of healthcare and nursing. This architecture goes beyond simple business automation, creating a new model of care: data-driven predictive intervention. The system achieves the crucial technological effect of providing maximum quality care with limited human resources. This technology is extremely useful in specific application areas such as health management infrastructure in smart cities or next-generation highly integrated healthcare facilities.

[0112] (Implementation steps) Step 1: Diverse Collection of Health Data The system automatically acquires various health-related data from wearable devices such as smartwatches, activity meters, blood pressure monitors, thermometers, pulse oximeters, and blood glucose meters, as well as from health management applications, electronic medical records from regular checkups, and medication management applications installed on smartphones and tablets. This data includes heart rate, blood pressure, body temperature, steps, sleep duration, weight, blood glucose levels, oxygen saturation, calories burned, resting heart rate, exercise intensity, wake-up and bedtime times, past medical history, family medical history, medication history, allergy information, and results of regular checkups, and sends this data to the server. For example, it obtains daily morning blood pressure measurements from a smartwatch via Bluetooth, weekly step counts from an activity meter, monthly weight changes from a scale via Wi-Fi, medication status from a medication management application, and automatically links to electronic medical records from medical institutions for regular checkups. Specifically, in step one, the system executes an edge gateway module that comprehensively receives bio-signal streams asynchronously transmitted from various IoT devices. This module simultaneously terminates multiple communication protocols, including Bluetooth Low Energy and Wi-Fi standards. This module extracts the payload from the received data packets. It then converts the extracted data into a time-series database write format. The input data to this module consists of binary sensor data packets sent from various devices. The dimensionality of the input data depends on the type of sensor used in the device; it can be a one-dimensional or multi-dimensional array. Specific examples of input data include numerical arrays of ECG waveforms sampled at 100 Hz or triaxial acceleration data at one-second intervals. This module applies a low-pass filter to the received data. It removes high-frequency noise. This module constructs an encrypted tunnel for securely forwarding the noise-removed data to the cloud server. The output data from this module is in the form of time-series data records structured in JSON format. Specific examples of output data include pairs of timestamps and heart rate values, or metadata containing the media access control address of the measuring instrument. This system sends this data to a distributed message queue system in the cloud. This system creates an environment where subsequent processing can consume data asynchronously and scalably. This structure prevents data loss even in the face of changes in communication bandwidth or temporary device interruptions. This system delivers a significant technological boost to the robustness and reliability of large-scale data collection. This technology is extremely useful in specific application areas such as health monitoring projects involving tens of thousands of citizens or real-time vital sign monitoring systems for hospitalized patients in large-scale hospitals.

[0113] Step 2: Comprehensive collection of daily activities, communication, and nursing records. Using the user's terminal's GPS function or accelerometer, calendar applications, activity tracking applications, voice input applications, image recognition applications, call history, messaging applications, SNS usage, voice recognition function, video call applications, email applications, camera applications, in-facility IoT sensors, or conversation recording systems, the system acquires data such as frequency of outings, distance traveled, amount of exercise, participation in hobby activities, household chores performed, sleep patterns, meal times, bathing times, toilet visits, TV viewing time, reading time, whether or not there is a nap, diversity of walking routes, number of calls, number of messages sent and received, SNS posting frequency, conversation length, topic diversity, presence of emotional expression, repetition of the same topic, response speed to questions, frequency of topic switching, participation in entertainment activities, and records of accidents or abnormal behavior, and sends this data to the server. Specifically, in step two, the system executes a sensor fusion algorithm to quantify the user's movement trajectory or interaction with the environment. The system obtains latitude and longitude coordinates from the GPS receiver built into the smartphone. The system obtains three-dimensional motion vectors from the accelerometer and gyroscope sensors. The system inputs this spatial data into a Kalman filter. This system corrects for sensor measurement errors and estimates high-precision movement trajectories. The input data is a heterogeneous, mixed time-series numerical stream obtained from multiple sensors. The dimensionality of the input data is an eight-dimensional tensor combining two dimensions of location coordinates and six dimensions of motion vectors. Specific examples of the input data include floating-point values ​​of latitude and longitude or continuous values ​​of angular velocity updated per second. The estimated movement trajectory is input into a spatial clustering algorithm. The system extracts frequent user locations or movement patterns. Unstructured data obtained from voice input applications or call logs is input into the natural language processing module. The system measures the length of speech or the duration of silence. The output data from this system is a multi-dimensional feature vector integrating action and communication features. Specific examples of the output data include values ​​representing the total distance traveled in a day or values ​​in milliseconds representing the average response time in a conversation. The system sends these feature vectors to a cloud server via a secure communication channel. Through this structure, the system continuously captures subtle changes in daily life in the background without relying on conscious user input. This system effectively generates objective and comprehensive life logs automatically. This technology is extremely useful in specific application areas such as anomaly detection systems for elderly people living alone or behavior change monitoring systems for patients with mental illness.

[0114] Step 3: Data Integration and Preprocessing On the server, health data, daily activity data, communication data, and nursing record data are integrated for each user to generate an integrated dataset managed by time series. The system also performs data correction for missing or outlier values, anonymization for privacy protection, and data normalization or standardization. For example, it deduplicates step count or sleep data obtained from multiple devices or applications, automatically detects and corrects outliers, and converts all data into a common time series format for management. Specifically, in step three, the system executes a data orchestration engine that controls the extraction, transformation, and loading pipeline. This engine temporarily stores various data streams arriving asynchronously from a distributed message queue in a buffer in memory. The engine combines multiple data streams using timestamps as keys. The engine executes a machine learning model using a bidirectional long short-term memory network for missing value imputation. The data input to this model is a multivariate time series matrix arranged along the time axis. The dimensionality of the data input to this model is equal to several hundred dimensions of the collected sensor data. As a specific example of the data input to this model, there may be a partial array of missing sleep depth data from one week or irregularly measured blood pressure values. This model uses contextual information from both the past and future to infer the most accurate value of the missing part. The engine executes the Isolation Forest algorithm for outlier detection. The engine identifies data points that significantly deviate from the normal data distribution and applies a smoothing filter. The output data from this engine is in the form of a fully imputed and normalized dense multidimensional tensor. As a specific example of the output data from this engine, there exists a 128-dimensional feature tensor uniformly sampled on an hourly basis. This engine stores this tensor in a columnar database. The engine optimizes the response speed to subsequent analytical queries. Through this structure, the system automatically transforms dirty data containing noise or missing data specific to the real-world environment into high-quality clean data that can be directly processed by the artificial intelligence model. This system effectively reduces the computational cost required for data preprocessing and improves the overall processing efficiency of the system. This technology is extremely useful in specific application areas such as the construction of large-scale medical data warehouses or the integration platform of nursing records from multiple facilities.

[0115] Step 4: Assessment of cognitive and physical functions (utilizing generative AI) The system analyzes and integrates datasets, utilizing generative AI to calculate cognitive and physical function scores from users' daily conversations, behavior patterns, health data, and nursing records. The generative AI is provided with prompts such as "Extract signs of changes or declines in cognitive or physical function from the user's conversation records, behavior records, health data, and nursing records over the past month, and score them," and "Detect patterns such as repetitive conversations or delayed responses, reduced vocabulary, decreased walking speed, increased number of falls, and decreased food intake, and assess the risks." Specifically, in step four, the system executes a multimodal inference engine centered on a large language model. This engine reads the multidimensional tensors generated in the data integration department. It executes a template generator to convert numerical data into natural language descriptions. The engine combines the converted text with free-form descriptive text from nursing records. This combined text is then input into the large language model as a prompt. The data input to this model is an array of lexical indices segmented from the text using byte-pair encoding. The maximum dimension of the data input to this model is 8,192 lexical units. As specific examples of data input to this model, there are text terms indicating a 10% decrease in walking speed compared to the previous month, or text terms indicating the same question was repeated three times in a conversation. This model uses hundreds of layers of Transformer blocks to compute complex self-attention weights between input terms. The model extracts signs of cognitive and physical function decline while comparing these values ​​to pre-learned medical knowledge. The output data from this model is a structured JSON object containing assessment scores for each function and the accompanying text. As specific examples of output data from this model, there is a numerical value of a cognitive function score of 65, or text data suggesting a decline in short-term memory. The engine parses the output JSON object and appends the scores to a time-series database. Through this structure, the system achieves a high degree of semantic parsing, comprehensively interpreting both numerical and unstructured text data. This system demonstrates the technical effectiveness of automatically and instantaneously performing comprehensive assessments approaching those of human experts. This technology is extremely useful in specific application areas such as early detection screening tools for dementia or systems for measuring the effectiveness of rehabilitation.

[0116] Step 5: Risk Prediction (Utilizing Generative AI) Using a generative AI-based machine learning model, this study comprehensively analyzes family or past medical history, changes in cognitive and physical scores, fall history, medication status, frequency of social interaction, and changes in eating or sleeping patterns to individually predict the risk of dementia or physical function decline. The generative AI is then tasked with "predicting the risk of dementia or physical function decline based on the user's integrated dataset and identifying the risk factors." “Identify lifestyle habits or behavior patterns that contribute to increased risk and create improvement proposals.” The system then executes a risk prediction pipeline in step five, linking a deep learning-based survival analysis model with a large language model. This system combines the time-series shifts in assessment scores calculated in step four with static features such as family history or medication status. The combined data is then input into a multilayer perceptron network. The input data is a one-dimensional tensor connecting continuous values ​​and categorical variables. The data has 512 dimensions. Specific examples of the input data include numerical values ​​representing the slope of cognitive function scores over the past six months or unique heat vectors representing the use of specific antihypertensive medications. The network uses a non-linear activation function to learn the complex interactions between features. The network outputs the probability of developing dementia or falling within the next year. The calculated probability values ​​and high-contribution features are then input into the large language model. The input data is a prompt string containing a list of risk probabilities and features. Based on these values, the model generates a logical explanation of the risk factors and specific improvement proposals. The output data from this model is in the form of risk report text described in natural language. Specific examples of the model's output data include statements such as increased fall risk due to the use of sleeping pills and increased nighttime urination, or improvement proposals such as installing sensor lights at the bedside. The system stores the generated reports in a database. Through this structure, the system achieves both high-precision probability prediction and the generation of descriptions in human-understandable natural language. This system effectively increases the acceptance of artificial intelligence judgments by healthcare professionals and families, prompting rapid intervention. This technology is extremely useful in specific application areas such as accident prevention systems in nursing facilities or high-risk patient retrieval systems in regional healthcare networks.

[0117] Step 6: Generate individually optimized care plans (utilizing generative AI) Based on risk prediction results or the user's interests, lifestyle, family structure, and past activity history, the generative AI is given prompts such as "Create a one-week rehabilitation menu, life support suggestions, recreational activity proposals, and feedback for family and staff, tailored to the user's risks and interests," and "Generate memory training content using the user's photos or memories," automatically generating an individually optimized care plan. Specifically, in step six, the system executes a planning algorithm using constrained language generation technology. The system obtains the risk report generated in step five and personal information extracted from the user's interest database. The system integrates this information and embeds it into the prompt template. The system inputs the completed prompts into a large language model. The data input into this model is an array of text lexical units containing the user's current state, the target health status, and constraints on available resources. The data input into this model has thousands of lexical units. Specific examples of data input into this model include text indicating a constraint of avoiding strenuous exercise due to a high risk of falls, or personal information indicating a past enjoyment of gardening. The model generates specific activity plans that satisfy the constraints while also attracting the user's attention through an autoregressive decoding process. This model uses a bundle search algorithm to explore sequences of plans with the highest likelihood and greatest diversity. The output data from this model is structured schedule data containing date and time, activity name, required time, and necessary preparations. As a specific example of the output data, there is a schedule for gardening work to be performed for 30 minutes starting at 2 PM on Tuesday, while seated, or a list of the necessary tools. The system registers the generated schedule data in the calendar system. Through this structure, the system automatically constructs fully personalized, executable care plans for each user from countless combinations. This system significantly improves the efficiency of care managers' plan creation operations and enhances the individualization of services. This technology is extremely useful in specific application areas such as automated scheduling systems for daily work in day care centers or self-care support applications for home-based caregivers.

[0118] Step 7: The AI-powered Nursing Counseling Department proposes, implements, and monitors (utilizing generative AI). Through tablets, smartphones, voice assistants, chatbots, etc., it suggests generated nursing plans, rehabilitation menus, life support suggestions, and provides feedback to users or caregivers, monitoring implementation and user reactions. It provides the generative AI with prompts such as "generating optimal encouragement, reminders, attention-grabbing suggestions, and action proposals based on user activity implementation, emotional expression, lifestyle rhythm, changes in physical condition, and feedback from family or caregivers," and "automatically updating nursing plans or rehabilitation menus based on user reactions or implementation." Specifically, in Step 7, the system executes a reinforcement learning-based dialogue manager to control the interactive AI agent. The system reads the nursing plan generated in Step 6 as the agent's action plan. The system acquires real-time facial expression or voice data from the user's camera or microphone. This data is input into a convolutional neural network for emotion recognition. The data input to this network is an image tensor of the face or a Mel-frequency cepstral coefficient matrix of the voice. The dimension of the data input to the network is a three-channel tensor of 256 pixels square in the case of images. Specific examples of data input to this network include image data representing a user's fatigued expression or low-pitched speech data. The network digitizes the current user's emotional state or fatigue level. This system inputs this emotional state and the pre-defined care plan into a large language model. The model generates the most suitable conversational text for the current situation. The output data from this model takes the form of a text string for vocalization and a set of parameters controlling the prosody of the synthesized speech. Specific examples of output data from this model include text such as "You seem a little tired, don't push yourself, how about just five minutes of stretching," or parameters setting a slower speech rate. This system uses a text-to-speech synthesis engine to generate natural speech and outputs it from the terminal's speaker. The system uses sensors to detect whether the user has implemented the proposal and uses this result as a reward for reinforcement learning to update the model. Through this structure, the system achieves dynamic and empathetic interaction that closely reflects the user's immediate state. This system effectively promotes user behavior change and dramatically improves the persistence rate of plans. This technology is extremely useful in specific applications such as communication robots for the elderly or smart speaker applications aimed at improving medication adherence.

[0119] Step 8: Regular Reassessment and Feedback (Utilizing Generative AI) Regularly reassess cognitive or physical function, using generative AI to analyze changes in scores or risk levels. Provide prompts such as "Summary the user's activity history and cognitive and physical function score shifts over one month, creating a report for the family" and "Summary risk factors and recommended intervention strategies for medical institutions," automatically generating optimal feedback or reports for families, medical institutions, and caregivers. Specifically, in Step 8, this system executes a document generation engine that summarizes long-term time-series data and generates reports corresponding to the subject. This system extracts accumulated cognitive function scores, physical function scores, and daily activity log data from the database over the past month. The extracted data is input into the statistical processing module to calculate statistical features such as mean, variance, and trend slope. This system inputs these statistical data and raw log data into a large-scale language model. The data input to this model is a long-text contextual prompt combining numerical lists and text logs. The dimension of the data input to the model reaches tens of thousands of tags, serving as the maximum context length for the model. Specific examples of data input to this model include numerical arrays representing step count trends over a month, or summary texts of conversations with the nursing counseling department. The model extracts key changes or noteworthy findings from the massive input data. It adjusts the tone and level of technical terminology in its output based on the reader's attributes (family, physician, nurse). The output data is presented as structured report text in Markdown format. Specific examples of output data include warm, jargon-free reports for families, or concise, itemized summaries of outliers and risk factors for physicians. The system publishes the generated reports on a portal and sends notifications to relevant personnel. Through this structure, the system automatically extracts intuitively understandable insights from massive datasets. This system enhances the quality of information sharing across multiple professions and strengthens teamwork in nursing. This technology is particularly useful in specific applications such as the automatic generation of discharge summaries in electronic medical record systems and the automatic creation of monthly reports in nursing facilities.

[0120] Step 9: External Collaboration and Emergency Response. The system collaborates with healthcare institutions, care providers, regional integrated support centers, families, social welfare personnel, care managers, pharmacists, nutritionists, rehabilitation staff, and regional event organizers. With the user's consent, cognitive function scores, risk levels, and activity history are sent to healthcare institutions to adjust care plans. The system automatically retrieves regional activity information and proposes it to the user. In emergencies such as falls or sudden cognitive decline, the system automatically notifies families or healthcare institutions. Specifically, in Step 9, the system executes an external collaboration gateway based on an event-driven asynchronous communication architecture. The system has a rule-based composite event processing engine that continuously monitors the user's status. The system inputs sensor data or assessment score streams into this engine. The data input to the engine is a real-time event message stream arriving in milliseconds. The dimension of the data input to the engine is a variable-length key-value pair depending on the event type. Specific examples of data input to the engine include numerical data representing a sudden impact detected by an accelerometer or an alarm flag indicating a heart rate below a threshold. The engine parses the combination patterns of multiple events and instantly determines the occurrence of an emergency. When the engine detects an emergency, the system immediately initiates an interruption process. The system references a pre-registered database of contact information for medical institutions or families. The system invokes an application programming interface (API) for emergency notification. Output data from the system is in the form of a POST request payload using the Hypertext Transfer Protocol (HTTP). Specific examples of output data include JSON data containing the user's current latitude and longitude, or vital sign updates over the last five minutes. The system encrypts this data and sends it to an external system, implementing retransmission control until confirmation of receipt is received. Through this structure, the system detects life-threatening events with millisecond-level latency and effectively communicates them to relevant authorities. This system achieves the crucial technical effect of improving life-saving rates and preventing severe illness. This technology is extremely useful in specific applications such as emergency notification systems for elderly people living alone or wide-area care networks in smart cities.

[0121] (specific use case) For example, when the system of this invention is implemented in a nursing facility, users wear smartwatches or wearable devices such as activity meters, blood pressure monitors, thermometers, pulse oximeters, and blood glucose meters to automatically acquire health data such as heart rate, blood pressure, body temperature, steps, sleep time, weight, blood glucose levels, oxygen saturation, calories burned, resting heart rate, exercise intensity, and wake-up and bedtime. Furthermore, nursing staff use tablet terminals or nursing record systems to record in detail the amount or content of meals, the presence or time of bathing, the frequency or condition of bowel movements, the use of walking aids during movement, the date, time, and condition of falls or tumbles, the frequency or content of forgetfulness, the content or response of conversations, participation in recreational activities or responses, medication administration or forgetting to take medication, vital sign measurements, rehabilitation implementation or achievement, details of accidents or abnormal behavior, etc. In addition, data such as frequency of outings, distance traveled, amount of exercise, participation in hobby activities, household chores, sleep patterns, meal times, bath times, toilet visits, TV viewing time, reading time, naps, variety of walking routes, number of calls, number of messages sent and received, frequency of SNS posts, conversation length, topic variety, emotional expression, repetition of the same topic, response speed to questions, frequency of topic switching, participation in recreational activities, and records of accidents or unusual behavior are also collected using user terminals' GPS or accelerometer sensors, calendar applications, activity tracking applications, voice input applications, image recognition applications, call logs, call history, SNS usage, SNS usage, SNS activity, voice recognition, video calling applications, email applications, camera applications, and IoT sensors or conversation recording systems within the facility. Specifically, this system constructs a data collection foundation that combines a high-density sensor network and edge computing within the enclosed space of a nursing facility. The system continuously collects biometric data from various wearable devices worn by users using low-power wireless communication standards. The system uses multiple receiving antennas configured within the facility to measure the radio wave intensity from the devices. This system employs a tri-point positioning algorithm to determine the user's indoor location with centimeter-level accuracy. Input data consists of arrays of radio wave intensity values ​​received by each antenna, or packets of vital sign data transmitted from the device. The dimension of the input data is a multidimensional vector proportional to the product of the number of antennas and sensor channels. Specific examples of input data include radio wave intensity values ​​of -60 dBm or blood oxygen saturation values ​​of 98%. This system reduces these datasets to an edge server, timestamps them, and stores them in a time-series database. Similarly, the system integrates electronically recorded data input by caregivers or skeletal estimation coordinate data extracted from camera images within the facility. Output data from this system is a multidimensional life log tensor containing each user's spatial and temporal movement trajectory.As specific examples of output data from this system, there are numerical values ​​representing the time spent in the cafeteria, or time-series arrays representing proximity to others. Through this structure, the system captures all events within the facility as digital data without omission. This system achieves the technical effect of accurately recording even changes in state at night or in individual rooms, which are invisible to the human eye. This technology is extremely useful in specific applications such as automating caregiving operations in large specialized nursing homes for the elderly, or event prevention systems in dementia wards of hospitals.

[0122] This diverse data is integrated and managed by the Data Integration Department, categorized by user and time series, and undergoes processes such as correction of missing or outlier values, anonymization, normalization, and standardization. In the Cognitive and Physical Function Assessment Department, generative AI is utilized to provide prompts such as "Extract signs of changes or declines in cognitive or physical function from the user's conversation records, activity records, health data, and care records over the past month, and assign a score," and "Detect patterns such as repetitive conversations or delayed responses, reduced vocabulary, decreased walking speed, increased number of falls, and reduced food intake, and assess the risks." Generative AI calculates cognitive and physical function scores based on information from various aspects, including the content or frequency of conversations, reports of forgetfulness, walking or movement patterns, records of falls or accidents, eating or excretion status, frequency of social interaction, and participation in recreational activities. Specifically, this system inputs the collected multidimensional life log tensors into a large-scale language model in the cloud to perform advanced semantic reasoning. In the Data Integration Department, this system executes an encoder network that transforms heterogeneous mixed data into a common latent space vector. After imputing missing values ​​or removing noise, the system passes the data to the prompt generation module. This module maps numerical or categorical data to a natural language context. The generated prompts are then fed into a large Transformer-based language model with billions of parameters. The data input to this model consists of lengthy, labeled sequences of text containing the user's activity history, vital sign progression, and care records. The dimension of the input data ranges from thousands to tens of thousands of labels. Specific examples of the input data include text labels indicating that nighttime wandering occurs three times a week, or text labels indicating that meal intake has decreased to half of normal. The model uses a self-attention mechanism to calculate the hidden causal relationships between these events. The model does not simply detect anomalies but performs a comprehensive scoring based on medical knowledge. The output data from this model is in the form of continuous scores from 0 to 100 for various domains of cognitive and physical function, along with textual data explaining the basis of these calculations. Specific examples of the output data from this model include a cognitive score of 50 indicating progression of disorientation, or a physical score of 60 indicating decreased lower limb muscle strength. This system visualizes these scores on a dashboard and provides feedback to staff. This structure allows for the early detection of complex functional declines that are difficult to discern from fragmented data. The system enables proactive care interventions to prevent the progression of the disease. This technology is particularly useful in specific applications such as measurement systems for the effectiveness of preventative care programs and health management platforms in homes for older adults.

[0123] In the Risk Prediction Department, a machine learning model based on generative AI is used to comprehensively analyze family medical history or past medical history, changes in cognitive and physical scores, fall history, medication status, frequency of social interaction, and changes in eating or sleeping patterns. This provides the generative AI with prompts such as "Predict the risk of dementia or physical function decline based on the user's integrated dataset and explain the risk factors" and "Extract lifestyle habits or behavior patterns that contribute to increased risk and create improvement proposals," determining the risk level (low, medium, high) and outputting risk factors or improvement points in natural language. Specifically, this Risk Prediction Department functions primarily based on a large-scale language model with a Transformer architecture possessing a self-attention mechanism. As input data, this department converts the user's time-series data into multi-dimensional numerical vectors. It also converts static data such as family medical history or past medical history into classification vectors using one-hot encoding. Finally, it normalizes dynamic data such as changes in cognitive and physical scores or sleep patterns as time-series arrays for each time step. This risk prediction unit combines these vector groups to generate, for example, a 512-dimensional integrated feature tensor. This integrated feature tensor is then input into the embedding layer of a large language model. The risk prediction unit segments the aforementioned prompt text into natural language token columns. These token columns, along with positional encodings, are input into the model. The large language model calculates the attention weights between the integrated feature tensor and the prompt's token columns. Through nonlinear operations in high-dimensional space, the large language model outputs a probability distribution representing the risk level. The output data from this large language model is represented by the output values ​​of the Softmax function for low, medium, and high risk categories. For example, the output data contains probability distribution vectors for low risk (0.1), medium risk (0.2), and high risk (0.7). Simultaneously, the large language model autoregressively generates natural language text token columns describing the risk factors. For example, the output text contains the string "Reduction in sleep time and decrease in social interaction frequency over the past 3 months are extracted as major risk factors for cognitive decline." The risk prediction unit determines the corresponding risk level when the maximum value of the output probability distribution exceeds a predetermined threshold. The risk prediction department sends the determined risk level to the subsequent care plan generation department. This processing method does not rely on human intuition but mathematically analyzes the complex correlations between multiple variables. Through this method, the system detects subtle risk signs at an early stage. This system significantly improves prediction accuracy compared to existing methods. The system pre-optimizes weight parameters using past user datasets through supervised learning of the cross-entropy error function. The system uses gradient descent to minimize the model's inference error. As a result, this technology reduces the parsing time of massive amounts of medical and nursing data, achieving high-precision risk prediction while minimizing computational resource consumption.As a specific application area, this system is used in preventive intervention systems in nursing facilities or remote monitoring systems in home care.

[0124] In the care plan generation department, based on risk prediction results or the user's interests (such as music, sports, gardening, cooking, reading, handicrafts, Go, Shogi, karaoke, pets, travel, etc.), living conditions, family structure, and past activity history, the generative AI is provided with prompts such as "create a one-week rehabilitation menu, life support suggestions, entertainment activity proposals, and feedback for family and staff, tailored to the user's risks and interests," and "generate memory training content using the user's photos or memories," automatically generating individually optimized care plans. For example, in high-risk situations, it strongly recommends medical institution visits or specialist consultations; it proposes music therapy or singing entertainment activities to users who like music; and it proposes individual rehabilitation or walking programs to users who like sports. In addition, it automatically generates feedback for family or caregivers such as "recent changes in cognitive or physical function," "key points to pay attention to in the future," and "examples of conversation in daily life." Specifically, this care plan generation department has a generative neural network capable of processing multimodal input data. As input data, this care plan generation department obtains the probability distribution vector of the risk level received from the risk prediction department. This nursing plan generation department uses natural language processing (NLP) technology to convert text data representing a user's interests or lifestyle into feature vectors. It also obtains past activity history as a time-series action log matrix. This department integrates this heterogeneous data to construct, for example, a 1024-dimensional context tensor. This context tensor, along with the aforementioned prompts, is then fed into an encoder-decoder structure of a large language model. Based on the input context tensor, the large language model performs a mapping of features in the latent space. As output data, this large language model generates structured nursing plan data. For example, the output data includes identifiers for daily rehabilitation menus, category codes for recommended recreational activities, and text strings containing family-oriented suggestions as elements. The large language model also converts the user's photo image data into an image feature tensor using a convolutional neural network. This large language model combines the image feature tensor and text feature vectors through a cross-attention mechanism. From this combined data, the large language model outputs personalized quiz-style text for memory training. For example, the output data includes the question "Where was this photo taken?" and labels for the correct answers. The nursing plan generation department parses the generated structured data and transforms it into a set of commands that can be interpreted by the subsequent AI-powered nursing guidance department. This processing method does not rely on rule-based conditional branching but rather leverages the generalization capabilities of deep learning models. Through this method, the system dynamically generates optimal plans for countless combinations of existing user attributes. The system uses reinforcement learning algorithms, employing user performance on past plans or satisfaction scores as rewards to fine-tune the model. The system uses policy gradient descent to update weights, generating more effective plans.As a result of this technology, the system eliminates the need for plan creation based on the personal experience of nursing managers. The system reduces the computational processing time required for plan creation to milliseconds. Through a data-driven approach, the system improves user compliance. Specific applications of this system include the automated creation of daily activity plans in day care centers and access to nursing plan creation support systems in nursing stations.

[0125] The AI-powered Nursing Counseling Department uses tablets, smartphones, voice assistants, and chatbots to suggest nursing plans, rehabilitation menus, life support recommendations, and feedback to users or caregivers, while monitoring implementation and user reactions. For example, it provides specific action proposals or raises awareness such as "Let's do 10 minutes of music therapy today," "Let's take a walk and stretch this afternoon," "We've been talking less lately, try video calling your family," and "Falls have increased, consider using walking aids." When users or staff input implementation details or reactions, this information is again accumulated in the data integration department and used in future assessments or plan generation. Furthermore, generative AI is used to analyze users' activity implementation, emotional expression, lifestyle rhythm, changes in physical condition, and feedback from family or caregivers in real time. This provides prompts such as "generating optimal encouragement, reminders, attention-grabbing suggestions, and action proposals based on the user's activity implementation, emotional expression, lifestyle rhythm, changes in physical condition, and feedback from family or caregivers," and "automatically updating the care plan or rehabilitation menu based on the user's response or implementation." Specifically, this AI-powered nursing guidance system is installed as an edge-cloud collaborative system integrating a real-time stream processing engine and an interactive large-scale language model. As input data, this AI-powered nursing guidance system receives multimedia data acquired from the user's terminal's camera or microphone. It converts the speech data into acoustic feature matrices such as Mel-frequency cepstral coefficients. It extracts facial landmark coordinates from image data to generate numerical vectors representing emotional states. This AI-powered nursing guidance department acquires binary flags representing activity implementation status, or biosensor data such as heart rate, as time-series arrays. This AI-powered nursing guidance department approximates these real-time datasets into, for example, a 256-dimensional dynamic state tensor. This dynamic state tensor, along with the aforementioned prompts, is input into a large-scale language model. This large-scale language model contains layers of recurrent neural networks or long short-term memory networks, maintaining a hidden state of past dialogue history. This large-scale language model parses the input dynamic state tensor as output data, outputting the text message to be prompted to the user, marked language data for speech synthesis, and a timestamp indicating the timing of the prompt. For example, the output data includes the text string "Great rhythm!", a speech synthesis instruction containing parameters of joyful emotion, and a flag indicating immediate execution. If the feature quantity representing user fatigue exceeds a predetermined threshold, this large-scale language model outputs an update command to reduce the intensity of the rehabilitation menu. This AI-powered nursing guidance department feeds back the output update command to the data integration department and the nursing plan generation department. This processing method differs from assessments with delays associated with human observation, detecting state changes with low latency measured in milliseconds.This system employs this processing method to immediately and adaptively intervene in the user's subtle reactions. It utilizes the Q-learning algorithm from reinforcement learning to optimize dialogue strategies and maximize the user's positive responses. As a result, the system reduces network traffic while enabling rapid inference processing at the edge device. Furthermore, the system simultaneously maintains user motivation and ensures security. Specific applications of this system include interactive care robots for elderly people living alone, and virtual assistant applications in rehabilitation facilities.

[0126] In addition, regular reassessments of cognitive or physical function are conducted, using generative AI to analyze changes in scores or risk levels, providing prompts such as "summarize the user's activity history and cognitive and physical function score shifts over the past month, creating a report for the family" and "summarize risk factors and recommend intervention strategies for medical institutions," automatically generating optimal feedback or reports for families, medical institutions, and caregivers. In cases of a sharp decline in cognitive or physical function, or in the event of a fall or accident, the generative AI automatically summarizes the content and sends alerts such as "recommend consulting a specialist" and "please strengthen fall prevention procedures." Specifically, this system has a periodic assessment module based on a batch processing architecture. This periodic assessment module uses time-series datasets of accumulated activity history from the past month as input data. This periodic assessment module constructs the shifts in cognitive and physical function scores into a multi-dimensional time-series tensor. This periodic assessment module applies a one-dimensional convolutional neural network to this time-series tensor to extract feature vectors representing long-term trends. This periodic assessment module inputs the extracted feature vectors and the aforementioned prompts into a large-scale language model dedicated to natural language generation. This large-scale language model sets the input feature vectors as the initial state of the decoder network. As output data, this large-scale language model outputs structured documents optimized for each recipient. For example, output data for families includes text strings replacing technical terms with colloquial expressions, and arrays of coordinate data used to depict graphs representing score progressions. Output data for healthcare institutions includes a list of risk factors based on the International Classification of Diseases (ICD) codes, and numerical data such as p-values ​​representing statistically significant differences. This periodic evaluation module calculates the first or second derivative of the input time-series tensor. If the calculated derivative exceeds a pre-set safety threshold, this periodic evaluation module determines a sharp functional degradation. This periodic evaluation module, as an anomaly detection algorithm, performs calculations of reconstruction errors using an autoencoder. If the reconstruction error exceeds a threshold, this periodic evaluation module generates short message data for immediate alerts. This periodic evaluation module sends the generated alert data as push notifications to designated terminals via a communication interface. This processing method completely replaces manual data aggregation or report creation tasks performed by humans. Through this processing method, this system provides consistency evaluation based on objective data. This system employs fine-tuning techniques from natural language processing to adjust model weights to mimic the descriptive style of medical experts. As a result, the system reduces the human resources required for report creation and prevents delays in information sharing. This system also enables the prevention of severe illness through early detection of abnormalities. Specifically, this system can be applied to multi-professional collaboration platforms within regional integrated nursing systems or as an automated summary generation system linked to electronic medical record systems.

[0127] In summary, the system of this invention collects and integrates various user data in real time and comprehensively at the nursing site. It utilizes generative AI to automate and enhance cognitive and physical function assessment, risk prediction, individualized optimized care plan generation, continuous support and feedback from an AI-based nursing coaching department, early anomaly detection, and medical collaboration. This comprehensively improves the quality of care, supports user self-reliance, reduces the burden on families and staff, suppresses medical expenses, and extends healthy lifespan. Specifically, this system functions as a comprehensive cloud computing platform integrating multiple neural network models and a distributed database. As input data, the system continuously receives unstructured and structured data streams from a cluster of edge devices. The received data is stored in a data lake and converted into normalized multidimensional tensors through a preprocessing pipeline. These tensors are routed to various inference endpoints for assessment, prediction, plan generation, and coaching. At each inference endpoint, the system drives a parallel computing cluster using GPUs. The system performs matrix multiplication and activation function operations at high speed based on large-scale language models or time series analysis models. As output data, the system writes probability distributions, text tag sequences, and control commands generated from each module into the integrated database. This system directly controls the user terminal's display or voice output hardware based on output control commands. This processing method goes beyond simple information digitization, forming a data-driven autonomous feedback loop. Through this method, the system simultaneously processes variables exceeding the limits of human cognition, deriving a globally optimal solution. The system employs a federated learning architecture, protecting user privacy while sharing only the model's weight gradients among multiple facilities. The system continuously updates the global model using these shared gradients. As a result, this technology maximizes the overall system throughput and minimizes data communication latency. The system ensures scalability and fault tolerance through dynamic allocation of hardware resources. Specific application areas for this system include national-scale healthcare data infrastructure systems and elderly support infrastructure in smart cities.

[0128] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the smart device 14. In the smart device 14, the control unit 46A causes the output device 40 to output the result of the specific processing. The microphone 38B acquires voice representing the user's input on the result of the specific processing. The control unit 46A sends the voice data representing the user's input acquired by the microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0129] Data generation model 58 is what is known as generative AI (Artificial Intelligence).

[0130] One example of data generation model 58 includes ChatGPT (registered trademark) (Internet search).<URL: https: / / openai.com / blog / chatgpt> Generative AI, etc.

[0131] The data generation model 58 is obtained by performing deep learning on a neural network. A prompt containing instructions is input to the data generation model 58, along with inference data such as speech data representing speech, text data representing text, and image data representing images (e.g., data from still images or moving images). The data generation model 58 infers the input inference data based on the instructions shown in the prompt and outputs the inference result in one or more data forms, such as speech data, text data, and image data. The data generation model 58 may include, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization. The specific processing unit 290 performs the aforementioned specific processing while using the data generation model 58. The data generation model 58 may also be a fine-tuned model to output inference results from prompts without instructions; in this case, the data generation model 58 can output inference results from prompts without instructions. The data processing device 12, etc., includes multiple data generation models 58, including AI other than generative AI. AI other than generative AI includes, but is not limited to, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes. Furthermore, AI can also act as an AI agent. Additionally, when the processing described above is performed by AI, this processing may be partially or entirely performed by AI, but is not limited to this example. Moreover, processing performed by AI, including generative AI, can be replaced by rule-based processing, and vice versa.

[0132] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart device 14, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart device 14. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart device 14 or external devices, and the smart device 14 acquires or collects information required for processing from the data processing device 12 or external devices.

[0133] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart device 14.

[0134] [Second Implementation] Figure 3 An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0135] like Figure 3 As shown, the data processing system 210 includes a data processing device 12 and smart glasses 214. One example of the data processing device 12 is a server.

[0136] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0137] The smart glasses 214 include a computer 36, a microphone 238, a speaker 240, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, and communication I / F 44 are also connected to the bus 52.

[0138] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0139] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0140] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0141] Figure 4 An example of the main functions of the data processing device 12 and the smart glasses 214 is shown. Figure 4 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0142] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0143] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. The emotion inference function using the emotion-specific model 59 (emotion-specific function) includes various inferences and predictions related to the user's emotions, but is not limited to this example. Furthermore, emotion inference and prediction also include, for example, emotion analysis (analysis).

[0144] In the smart glasses 214, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 operating as a control unit 46A based on the receive / output program 60 executed on the RAM 48. Furthermore, the smart glasses 214 may also have the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and use these models to perform the same processing as the specific processing unit 290.

[0145] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the smart glasses 214. In the following description, the data processing device 12 will be referred to as the "server" and the smart glasses 214 will be referred to as the "terminal".

[0146] (Example 1) Since the specific processing flow is the same as that in Embodiment 1 described in the first embodiment above, the description is omitted.

[0147] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0148] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the smart glasses 214. In the smart glasses 214, the control unit 46A causes the speaker 240 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0149] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart glasses 214, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart glasses 214. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart glasses 214 or external devices, and the smart glasses 214 acquires or collects information required for processing from the data processing device 12 or external devices.

[0150] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart glasses 214.

[0151] [Third Implementation] Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0152] like Figure 5 As shown, the data processing system 310 includes a data processing device 12 and a head-mounted terminal 314. An example of the data processing device 12 is a server.

[0153] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0154] The head-mounted terminal 314 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a display 343. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, display 343, and communication I / F 44 are also connected to the bus 52.

[0155] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0156] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0157] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0158] Figure 6 An example of the main functions of the data processing device 12 and the head-mounted terminal 314 is shown. Figure 6 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0159] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0160] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the special processing unit 290.

[0161] In the head-mounted terminal 314, the processor 46 performs the receiving and output processing.

[0162] The memory 50 stores a receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 operating the control unit 46A based on the receive / output program 60 executed on the RAM 48.

[0163] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the head-mounted terminal 314. In the following description, the data processing device 12 will be referred to as the "server" and the head-mounted terminal 314 will be referred to as the "terminal".

[0164] (Example 1) Since the specific processing flow is the same as that in Example 1 described in the first embodiment above, the description is omitted.

[0165] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0166] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the head-mounted terminal 314. In the head-mounted terminal 314, the control unit 46A causes the speaker 240 and the display 343 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0167] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the head-mounted terminal 314, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the head-mounted terminal 314. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the head-mounted terminal 314 or external devices, and the head-mounted terminal 314 acquires or collects information required for processing from the data processing device 12 or external devices.

[0168] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the head-mounted terminal 314.

[0169] [Fourth Implementation] Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0170] like Figure 7 As shown, the data processing system 410 includes a data processing device 12 and a robot 414. An example of the data processing device 12 is a server.

[0171] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0172] Robot 414 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a control object 443. Computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, control object 443, and communication I / F 44 are also connected to the bus 52.

[0173] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0174] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0175] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0176] The controlled object 443 includes a display device, LEDs for the eyes, and motors for driving the arms, hands, and feet. The posture and movements of the robot 414 are controlled by controlling the motors for the arms, hands, and feet. Some of the robot 414's emotions can be expressed by controlling these motors. In addition, facial expressions of the robot 414 can also be expressed by controlling the illumination state of the LEDs for the robot 414's eyes.

[0177] Figure 8 An example of the main functions of the data processing device 12 and the robot 414 is shown. For example... Figure 8 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0178] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0179] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the special processing unit 290.

[0180] In robot 414, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on RAM 48. The receive / output processing is implemented by the processor 46 operating the control unit 46A based on the receive / output program 60 executed on RAM 48.

[0181] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the robot 414. In the following description, the data processing device 12 will be referred to as the "server" and the robot 414 will be referred to as the "terminal".

[0182] (Example 1) Since the specific processing flow is the same as that in Example 1 described in the first embodiment above, the description is omitted.

[0183] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0184] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the robot 414. In the robot 414, the control unit 46A causes the speaker 240 and the controlled object 443 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0185] Furthermore, the processing performed by the data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the robot 414, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the robot 414. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the robot 414 or external devices, and the robot 414 acquires or collects information required for processing from the data processing device 12 or external devices.

[0186] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the robot 414.

[0187] Furthermore, the emotion-specific model 59, serving as an emotion engine, can determine the user's emotion based on a specific mapping. Specifically, the emotion-specific model 59 can determine the user's emotion based on an emotion graph that serves as a specific mapping (see...). Figure 9 The system determines the user's emotions. In addition, the emotion-specific model 59 can also determine the robot's emotions in the same way, and the specific processing unit 290 can also perform specific processing using the robot's emotions.

[0188] Figure 9This is a diagram representing an emotion map 400 that maps various emotions. In the emotion map 400, emotions are arranged radially from the center in concentric circles. The closer to the center of the concentric circles, the more primitive the emotion is. Further out on the concentric circles, emotions are arranged representing states or actions arising from mood. Emotion is a concept that includes both feelings and mental states. To the left of the concentric circles, emotions generated by reactions occurring in the brain are arranged roughly. To the right of the concentric circles, emotions guided by situational judgments are arranged roughly. Above and below the concentric circles, emotions generated by reactions occurring in the brain and guided by situational judgments are arranged roughly. Furthermore, the emotion of "pleasure" is arranged above the concentric circles, and the emotion of "unpleasantness" is arranged below. Thus, in the emotion map 400, various emotions are mapped according to the structure of emotion generation, while easily generated emotions are mapped nearby.

[0189] These emotions are distributed at the 3 o'clock position on the Emotion Chart 400, and usually fluctuate between peace and unease. In the right half of the Emotion Chart 400, because situational awareness is more dominant than internal feelings, it gives a sense of calm.

[0190] The inner side of the emotion diagram 400 represents the mind, and the outer side of the emotion diagram 400 represents actions. Therefore, the further you go to the outer side of the emotion diagram 400, the more the emotion can be seen (manifested in actions).

[0191] Here, human emotions are based on a balance of various factors such as posture and blood sugar levels. When these balances deviate from the ideal, it indicates unhappiness; when they approach the ideal, it indicates pleasure. In robots, cars, and motorcycles, emotions can also be created based on a balance of factors such as posture and remaining battery power. When these balances deviate from the ideal, it indicates unhappiness; when they approach the ideal, it indicates pleasure. Emotion maps can be generated, for example, based on Dr. Mitsuyoshi's emotion map (Research on Speech Emotion Recognition and Brain Physiological Signal Analysis Systems for Emotion, Tokushima University, Doctoral Dissertation: https: / / ci.nii.ac.jp / naid / 500000375379). In the left half of the emotion map, emotions belonging to the "Reaction" domain, where sensation is dominant, are arranged. Furthermore, in the right half of the emotion map, emotions belonging to the "Situation" domain, where situational cognition is dominant, are arranged.

[0192] The emotion map defines two types of emotions that promote learning. One is a negative emotion located near the middle of "repentance" or "reflection" on the situation side. That is, when the robot experiences negative emotions such as "I never want to feel this way again" or "I never want to be scolded again." The other is a positive emotion located near "desire" on the response side. That is, when the robot experiences positive feelings such as "wanting more" or "wanting to know more."

[0193] The emotion-specific model 59 feeds user input into a pre-learned neural network to obtain emotion values ​​representing each emotion shown in the emotion graph 400, and determines the user's emotion. This neural network is pre-learned based on multiple learning data sets that combine user input with emotion values ​​representing each emotion shown in the emotion graph 400. Furthermore, this neural network is learned to... Figure 10 As shown in sentiment graph 900, sentiment values ​​in nearby configurations are similar to each other. Figure 10 Examples show that multiple emotions such as "peace of mind", "stability", and "reassurance" have similar emotional values.

[0194] The above description primarily focuses on the functions of the data processing device 12 to illustrate the system disclosed herein. However, the system disclosed herein may not necessarily be implemented on a server. The system disclosed herein can be implemented as a general information processing system. For example, it can be implemented as a software program running on a personal computer, an application running on a smartphone, etc. The methods disclosed herein can be provided to users in the form of SaaS (Software as a Service).

[0195] In the above embodiments, an example of specific processing performed by a single computer 22 is given. However, the technology disclosed herein is not limited to this, and distributed processing of specific processes by multiple computers, including computer 22, is also possible. For example, the data generation model 58 may also be provided in an external device of the data processing apparatus 12, where data is generated based on input data.

[0196] In the above embodiments, an example of storing a specific processing program 56 in memory 32 is illustrated, but the technology disclosed herein is not limited thereto. For example, the specific processing program 56 may also be stored in a portable computer-readable non-transitory storage medium such as a USB (Universal Serial Bus) memory. The specific processing program 56 stored in the non-transitory storage medium is installed in the computer 22 of the data processing device 12. The processor 28 performs specific processing according to the specific processing program 56.

[0197] Alternatively, the specific processing program 56 can be stored in a storage device such as a server connected to the data processing device 12 via a network 54, and the specific processing program 56 can be downloaded and installed into the computer 22 upon request from the data processing device 12.

[0198] Furthermore, it is not necessary to store the entire specific process 56 in a storage device such as a server connected to the data processing device 12 via the network 54, nor is it necessary to store the entire specific process 56 in the memory 32; a portion of the specific process 56 may also be stored.

[0199] As a hardware resource for performing specific processing, various processors can be used. For example, a CPU can be listed as a processor; it is a general-purpose processor that functions as a hardware resource for performing specific processing by executing software, i.e., a program. Furthermore, as a processor, a dedicated circuit can be listed as a processor; it is a processor with a circuit structure specifically designed for performing specific processing, such as an FPGA (Field-Programmable Gate Array), a PLD (Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit).

[0200] Every processor has built-in or connected memory, and every processor performs specific processes by using memory.

[0201] The hardware resources for performing a specific process can consist of one of these various processors, or a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs, or a combination of a CPU and an FPGA). Furthermore, the hardware resources for performing a specific process can also be a single processor.

[0202] As an example of a single processor, the first type consists of a combination of one or more CPUs and software, which functions as a hardware resource to perform specific processing. The second type uses a processor, such as a System-on-a-chip...

Claims

1. A system comprising a health data acquisition department, an action data acquisition department, a communication data acquisition department, a data integration department, a cognitive function assessment department, a risk prediction department, an intervention plan generation department, and an AI tutoring department, characterized in that, The health data acquisition unit automatically acquires users' vital signs, past medical history, family medical history, and other health-related data from external services or wearable terminals; the action data acquisition unit acquires users' daily action patterns, frequency of forgetfulness, implementation of social activities, amount of exercise, frequency of going out, and other action data; the communication data acquisition unit acquires users' conversation frequency, SNS or messaging application usage, content or frequency of interpersonal communication, and other communication data; the data integration unit integrates the health data, action data, and communication data to generate an integrated dataset for each user; the cognitive function assessment unit uses generative AI to analyze users' daily conversation content or action patterns based on the integrated dataset and calculates cognitive function scores; the risk prediction unit uses generative AI to individually predict the risk of developing dementia based on the cognitive function scores and the integrated dataset; the intervention plan generation unit uses generative AI to automatically generate individually optimized prevention programs based on the risk prediction results and users' interests or lifestyles; and the AI ​​guidance unit suggests the prevention programs to users via voice or text, monitors users' implementation, conducts regular cognitive function assessments and feedback, detects minor changes in cognitive function early, and recommends consultation with specialized institutions when necessary.

2. The system as described in claim 1, characterized in that, The AI ​​coaching department uses generative AI to continuously analyze changes in users' implementation or motivation, providing encouragement or reminders at the best time or in the best way for each user, and reflecting changes in users' lifestyles or hobbies in real time to dynamically update the intervention plan.

3. The system as described in claim 1, characterized in that, The intervention plan generation department uses generative AI to comprehensively consider multiple factors such as the user's cognitive function level or interests, past activity history, social interaction status, health status, family medical history, etc., and automatically creates an individually optimized prevention program that combines multiple interventions such as brain training programs, lifestyle improvement suggestions, introductions to social participation activities, and recommendations for visits to specialized institutions. Furthermore, generative AI is used to customize the content or difficulty of brain training content for each user.

Images