AIGC-based old people health diagnosis and accompanying system and method

Abstract

The invention discloses an AIGC-based old people health diagnosis and accompanying system and method, and relates to the technical field of smart pension, and the method comprises the steps: collecting multi-source heterogeneous data of old people in real time through an intelligent sensor, carrying out the preprocessing, and generating a structured multi-modal time series data set; fusing the structured multi-modal time series data set, pre-stored life experience data and health archives, and constructing and dynamically updating a personal full-dimensional knowledge base; retrieving associated memory materials from the personal full-dimensional knowledge base according to an emotion demand recognition result, and generating multi-modal accompanying content through AIGC; and based on the health risk assessment result and the multi-modal accompanying content, constructing a structured interaction script, and cooperatively executing health diagnosis assistance and accompanying content through a cross-device scheduling mode. According to the invention, body and mind integration and active elderly accompanying with memory continuity are achieved.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care technology, and in particular to a health diagnosis and care system and method for the elderly based on AIGC. Background Technology

[0002] AIGC-based health diagnosis and companionship technologies for the elderly occupy a crucial position in today's smart elderly care field. Through multimodal sensor data fusion, personalized knowledge modeling, and generative content synthesis, these technologies aim to achieve dynamic monitoring of the elderly's physiological state and proactive response to their emotional needs. Current technical solutions typically combine wearable devices to collect physiological indicators, utilize environmental sensors to obtain residential parameters, and capture interactive behaviors through voice or video terminals. Based on preset rules, they then generate health tips or standardized companionship content. Some systems also attempt to incorporate life experience information to enhance interactive affinity. Overall, these technologies reflect a trend towards personalized and proactive elderly care, and are gradually incorporating pattern recognition technology to improve the understanding of user states.

[0003] In the field of AIGC-based health diagnosis and care for the elderly, traditional methods of health diagnosis and care for the elderly have two shortcomings: First, health diagnosis and emotional care functions are separated, and physiological risk warnings are not linked with emotional state recognition, resulting in a lack of synergy between physical and mental health in intervention content; Second, the generation of care content relies on general corpora or static templates, failing to deeply associate real-time emotional themes with high-emotional-value life events and memory materials in an individual's full-dimensional knowledge base, making it difficult to achieve resonant companionship with individual memory anchors. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AIGC-based method for health diagnosis and care for the elderly to address the problems of the disconnect between health diagnosis and emotional care, as well as the lack of deep integration of personal life experiences into the care content.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for health diagnosis and care for the elderly based on AIGC, comprising: collecting multi-source heterogeneous data of the elderly in real time through intelligent sensors, preprocessing the data to generate a structured multimodal time-series dataset; integrating the structured multimodal time-series dataset, pre-stored life experience data, and health records to construct and dynamically update a personal full-dimensional knowledge base; analyzing health risk status and emotional needs based on the structured multimodal time-series dataset and the personal full-dimensional knowledge base, generating health risk assessment results, and identifying emotional needs; retrieving associated memory materials from the personal full-dimensional knowledge base based on the emotional needs identification results, and generating multimodal care content through AIGC; constructing a structured interactive script based on the health risk assessment results and the multimodal care content, and collaboratively executing health diagnosis assistance and care content through cross-device scheduling.

[0007] As a preferred embodiment of the AIGC-based elderly health diagnosis and care method of the present invention, the multi-source heterogeneous data includes physiological data, behavioral data, environmental data and emotionally rich interactive data. The preprocessing includes data cleaning, noise reduction, calibration, and timing alignment.

[0008] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for fusing structured multimodal time-series datasets, pre-stored life experience data, and health records are as follows: The structured multimodal time-series dataset is time-aligned and semantically matched with pre-stored life experience data to obtain the life experience fusion result; The structured multimodal time-series dataset is compared and analyzed with pre-stored health records, and the difference between the current value of physiological parameters and the individual's health baseline index is calculated to obtain the health record fusion results.

[0009] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for constructing and dynamically updating a personal full-dimensional knowledge base are as follows: Based on the integration results of life experiences and health records, an experience map and memory association network are constructed to generate an initial personal full-dimensional knowledge base; Based on a structured multimodal time-series dataset, the initial personal full-dimensional knowledge base is dynamically updated to generate a dynamically updated personal full-dimensional knowledge base.

[0010] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for generating health risk assessment results are as follows: Based on structured multimodal time-series datasets and pre-stored health records, the deviation of the current health status from the individual's health baseline indicators is calculated to obtain health status deviation data. The three-tiered early warning system (green-yellow-red) is used to comprehensively assess the degree of abnormality in physiological parameters, changes in behavioral patterns, and medication adherence in health status deviation data, thereby generating health risk assessment results.

[0011] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for identifying emotional needs are as follows: We will perform keyword matching and sentiment dictionary analysis on emotionally rich interactive data and behavioral data, extract sentiment keywords, and identify the user's current sentiment theme. The system associates a user’s current emotional theme with important life events in their experience graph, and determines whether the user is currently in an emotional state of nostalgia, loneliness, or need for social support, generating an emotional need identification result.

[0012] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for generating multimodal care content are as follows: The emotional tags in the emotional needs identification results are semantically matched with the metadata tags of multimodal memory materials in the personal full-dimensional knowledge base, and candidate multimodal memory materials associated with event nodes in the experience graph are retrieved from the memory association network. Calculate the semantic similarity between sentiment tags and metadata tags of candidate multimodal memory materials, sort them according to similarity scores, and select the multimodal memory material with the highest score as the selected multimodal memory material; AIGC uses selected multimodal memory materials as contextual cues to generate multimodal companionship content.

[0013] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for constructing a structured interactive script based on health risk assessment results and multimodal care content are as follows: Based on the warning level in the health risk assessment results, corresponding health prompts are generated, and the maximum playback time of the care content is set. The health prompts are played before the multimodal care content begins, and health intervention scheduling instructions are generated. The health tips and multimodal companionship content are sequentially integrated according to their playback order and duration limit to obtain a structured interactive script.

[0014] As a preferred embodiment of the AIGC-based health diagnosis and care method for the elderly described in this invention, the steps for collaboratively executing health diagnosis assistance and care content through cross-device scheduling are as follows: Parse device instructions in structured interactive scripts, distribute different content to corresponding target terminal devices, and output the content-device mapping and distribution results; Based on the mapping and distribution results between content and devices, cross-device scheduling enables each target terminal device to start playback according to the interaction sequence in the structured interactive script, and maintain playback synchronization and timing alignment.

[0015] Secondly, this invention provides an AIGC-based health diagnosis and care system for the elderly, comprising: a data preprocessing module for real-time collection of multi-source heterogeneous data from the elderly via intelligent sensors, preprocessing the data to generate a structured multimodal time-series dataset; a knowledge fusion module for fusing the structured multimodal time-series dataset, pre-stored life experience data, and health records to construct and dynamically update a personal full-dimensional knowledge base; a state recognition module for analyzing health risk status and emotional need status based on the structured multimodal time-series dataset and the personal full-dimensional knowledge base, generating health risk assessment results, and identifying emotional needs; a content generation module for retrieving associated memory materials from the personal full-dimensional knowledge base based on the emotional need identification results, and generating multimodal care content through AIGC; and a collaborative execution module for constructing a structured interactive script based on the health risk assessment results and multimodal care content, and collaboratively executing health diagnosis assistance and care content through cross-device scheduling.

[0016] The beneficial effects of this invention are as follows: by constructing and dynamically updating a personal full-dimensional knowledge base, a unified modeling of experience graphs and memory association networks is achieved, supporting accurate reasoning of pattern recognition in the context of an individual's life course; by constructing a structured interactive script, health intervention and emotional companionship are realized through cross-modal synchronous execution at the content, temporal, and device levels, enabling pattern recognition not only to be used for state discrimination, but also to drive proactive care with memory continuity and physical-mind synergy, ultimately achieving a personalized elderly care closed loop that deeply integrates personal background. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an AIGC-based approach to health diagnosis and care for the elderly.

[0019] Figure 2 This is a schematic diagram of an AIGC-based health diagnosis and care system for the elderly.

[0020] Figure 3 A flowchart for building and updating a personal, comprehensive knowledge base.

[0021] Figure 4 A flowchart for dual-path recognition of health and emotional states.

[0022] Figure 5 A flowchart for AIGC content generation and cross-device execution. Detailed Implementation

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0026] Reference Figures 1-5 This is one embodiment of the present invention, which provides a method for health diagnosis and care for the elderly based on AIGC, including the following steps: S1: Collect multi-source heterogeneous data of the elderly in real time through smart sensors, perform preprocessing, and generate a structured multimodal time series dataset; S1.1: Multi-source heterogeneous data includes physiological data, behavioral data, environmental data, and emotionally rich interactive data; Specifically, physiological data of the elderly is collected through smart sensors, including heart rate, blood pressure, blood oxygen saturation, body temperature, and blood glucose levels; behavioral data of the elderly is collected through wearable devices and non-contact sensors, including daily activity trajectories, gait characteristics, sleep rhythm changes, and medication records; environmental data of the space in which the elderly are located is collected through environmental sensing devices, including indoor temperature, humidity, light intensity, and ambient noise levels; and emotionally rich interactive data of the elderly is collected through smart terminals with voice and video functions, including voice dialogue content, facial expression video streams, handwritten input text, and touch screen interaction logs. After completing the above collection, multi-source heterogeneous data containing physiological data, behavioral data, environmental data, and emotionally rich interactive data is obtained.

[0027] S1.2: Perform data cleaning, denoising, calibration, and time-series alignment on multi-source heterogeneous data to generate a structured multimodal time-series dataset.

[0028] Specifically, the process involves cleaning the multi-source heterogeneous data, removing duplicate records, and interpolating to complete missing physiological, behavioral, environmental, and emotionally rich interactive data. The cleaned data is then denoised using low-pass filtering for physiological data, moving average filtering for environmental data, trajectory smoothing for behavioral data, and speech denoising for emotionally rich interactive data. The denoised data is then calibrated to standardize the units and formats of data from various sensors. Finally, the calibrated data is time-series aligned to ensure that physiological, behavioral, environmental, and emotionally rich interactive data are aligned to the same sampling frequency using a unified timestamp. This process generates a structured multimodal time-series dataset.

[0029] It should be noted that low-pass filtering is a signal processing method used to retain low-frequency components in a signal below a set cutoff frequency, while attenuating or removing high-frequency noise components above the set frequency.

[0030] S2: Integrate structured multimodal time-series datasets, pre-stored life experience data, and health records to build and dynamically update a personal full-dimensional knowledge base; S2.1: Perform time alignment and semantic matching between the structured multimodal time-series dataset and the pre-stored life experience data to obtain the life experience fusion result; Specifically, the timestamps in the structured multimodal time-series dataset are aligned with the event times in the pre-stored life experience data, establishing a correspondence between the behavioral data and emotionally rich interactive data in the structured multimodal time-series dataset and the important life events in the life experience data on a unified timeline. The text content in the pre-stored life experience data is segmented to extract keywords and named entities, including people, places, occupations, historical events, and emotional descriptions. The extracted keywords and named entities are semantically compared with the emotionally rich interactive data in the structured multimodal time-series dataset. Based on keyword matching and semantic similarity calculation, the semantic similarity between the two is obtained. For matching results where the semantic similarity is higher than the effective semantic matching threshold, the behavioral segments in the structured multimodal time-series dataset are linked to specific events in the life experience data. The life experience fusion result is then obtained.

[0031] It should be noted that the pre-stored life experience data refers to structured or semi-structured information about the personal life course of the elderly person, which is pre-stored during the system initialization or user registration phase through family member input, user oral transcription, and historical file import. This information includes, but is not limited to: the time, place, relevant people, and event descriptions of important life events (such as joining the army, getting married, retiring, and relocating); professional background and work experience; hometown information and growth environment; hobbies and artistic preferences (such as favorite operas, music, and books); family relationships and social networks; and memory materials with emotional value (such as captions for old photos, letter content, and background stories of commemorative items). This data is stored in the form of text, tags, or metadata and is used for subsequent time alignment and semantic matching with the structured multimodal time-series dataset collected in real time, supporting the construction of a personal full-dimensional knowledge base.

[0032] All of the above content has been agreed to by the user and is used for legitimate purposes.

[0033] The effective threshold for semantic matching is defined based on the principle of optimal comprehensive accuracy and recall of the semantic association between the sentiment-rich interactive data and the pre-stored life experience data in the structured multimodal time series dataset. The value ranges from 0.7 to 0.85, with an optimal value of 0.75.

[0034] The life experience fusion result generated after time alignment and semantic matching is specifically represented as a set of structured associated records. Each record contains the following fields: event identifier (a unique ID corresponding to a significant life event in the pre-stored life experience data), event occurrence time (the specific date in the original record), time alignment interval (a 15-day time window consisting of 7 days before and after the event occurrence time), associated behavioral fragments (daily activity trajectories, gait features, sleep rhythm changes, and medication behavior records extracted from the structured multimodal time-series dataset within the 15-day time window), associated emotionally rich interaction fragments (voice-to-text transcripts, facial expression video metadata, handwritten input content, and touch screen interaction logs extracted within the 15-day time window), keyword and named entity set (characters, locations, occupations, historical events, and emotional descriptions extracted from the event description), semantic similarity score (a value calculated from keywords / named entities and emotionally rich interaction fragments), and valid matching markers (marked as valid when the semantic similarity score is ≥0.75). The structured set of associated records is directly used to construct the experience graph in subsequent processes.

[0035] S2.2: Compare and analyze the structured multimodal time-series dataset with the pre-stored health records, calculate the difference between the current value of physiological parameters and the individual's health baseline indicators, and obtain the health record fusion results; Specifically, the system retrieves the mean systolic blood pressure, mean diastolic blood pressure, median resting heart rate, 5th percentile of blood oxygen saturation, body temperature diurnal rhythm fitting curve, and historical average fasting blood glucose from pre-stored health records as individual health baseline indicators. It then extracts corresponding systolic blood pressure, diastolic blood pressure, heart rate, blood oxygen saturation, body temperature, and fasting blood glucose values ​​from a structured multimodal time-series dataset. Finally, it calculates the difference between the systolic blood pressure value and the mean systolic blood pressure over the past 30 days, the difference between the diastolic blood pressure value and the mean diastolic blood pressure over the past 30 days, and the difference between the heart rate value and the mean resting heart rate over the past 30 days. The differences in the following parameters are used to determine whether the abnormal physiological parameters exceed the medically defined abnormal thresholds: the difference between the median resting heart rate, the difference between the blood oxygen saturation value and the 5th percentile of blood oxygen saturation, the difference between the body temperature value and the predicted value of the body temperature diurnal rhythm fitting curve at the corresponding time, and the difference between the fasting blood glucose value and the historical average fasting blood glucose value. Based on these differences, it is determined whether they exceed the medically defined abnormal thresholds, thereby identifying abnormal physiological parameter events. The number of consecutive days that the above differences exceed the medically defined abnormal thresholds within three consecutive days is counted. Medication adherence is determined in conjunction with medication behavior records. The number of consecutive days of abnormal physiological parameters and the results of medication adherence determination are integrated into the health record fusion results.

[0036] It should be noted that the medically defined abnormal thresholds are preset based on the criteria for judging abnormal physiological parameters in the clinical medical guidelines for the elderly population, and the range of values ​​is as follows: systolic blood pressure difference ≥20 mmHg, diastolic blood pressure difference ≥10 mmHg, heart rate difference ≥15 beats / minute, blood oxygen saturation difference ≤−3 percentage points, body temperature difference ≥0.8℃, and fasting blood glucose difference ≥2 mmol / L.

[0037] S2.3: Based on the integration results of life experience and health record, construct an experience map and memory association network to generate an initial personal full-dimensional knowledge base; Specifically, based on the important life events, their times, locations, related people, and emotional descriptions contained in the life experience fusion results, entity and relationship recognition is performed on the text content to construct a directed graph structure with identity information as the root node and life experience events as child nodes, forming an experience graph. Based on the number of days of abnormal physiological parameters, medication adherence judgment results, and personal health baseline indicators contained in the health record fusion results, health status tags are generated and attached as attributes to the event nodes in the corresponding time intervals of the experience graph. Text descriptions of old photos, letter content, background stories of commemorative items, and information on artistic preferences are extracted from pre-stored life experience data, and each memory material is assigned a metadata tag, which includes event time, location, people, emotional type, and content type. The memory materials with metadata tags are indexed and linked with the event nodes in the experience graph to form an association between event nodes and multimodal memory materials. The experience graph and the association are then integrated so that each event node contains both temporal and semantic relationships and is associated with multimodal memory materials, generating an initial personal full-dimensional knowledge base.

[0038] S2.4: Based on a structured multimodal time-series dataset, dynamically update the initial personal full-dimensional knowledge base to generate a dynamically updated personal full-dimensional knowledge base.

[0039] Specifically, the process involves extracting new behavioral data and rich emotional interaction data from a structured multimodal time-series dataset; identifying new activity types and habit changes in the new behavioral data and adding corresponding event information as new nodes to the experience graph; analyzing whether the rich emotional interaction data mentions unrecorded life experience events, and if so, extracting the event time, location, people, and emotional descriptions to generate new event nodes and inserting them into the experience graph; calculating the difference between the new physiological data and the individual's health baseline indicators to form new health status labels, and associating them with event nodes in the corresponding time intervals in the experience graph; annotating the new multimodal memory materials with metadata and establishing index links with the event nodes in the experience graph that are closest in time; and generating a dynamically updated personal full-dimensional knowledge base.

[0040] S3: Based on a structured multimodal time-series dataset and a personal full-dimensional knowledge base, analyze health risk status and emotional needs status, generate health risk assessment results, and identify emotional needs; S3.1: Based on the structured multimodal time series dataset and pre-stored health records, calculate the degree of deviation between the current health status and the individual's health baseline indicators, and obtain health status deviation data; Specifically, the system extracts systolic blood pressure, diastolic blood pressure, heart rate, blood oxygen saturation, body temperature, and fasting blood glucose values ​​from a structured multimodal time-series dataset. It retrieves the mean systolic blood pressure, mean diastolic blood pressure, median resting heart rate, 5th percentile of blood oxygen saturation, predicted values ​​of the circadian rhythm of body temperature at corresponding times, and historical average fasting blood glucose from pre-stored health records. The system calculates the differences between systolic blood pressure and the mean systolic blood pressure over the past 30 days, the differences between diastolic blood pressure and the mean diastolic blood pressure over the past 30 days, the differences between heart rate and the median resting heart rate, the differences between blood oxygen saturation and the 5th percentile of blood oxygen saturation, the differences between body temperature and the predicted values ​​of the circadian rhythm of body temperature at corresponding times, and the differences between fasting blood glucose and the historical average fasting blood glucose. It determines whether each difference exceeds a medically defined abnormal threshold and counts the number of consecutive days of abnormality over three days. It also assesses medication adherence based on medication behavior records and obtains data on deviations from health status.

[0041] It should be noted that the pre-stored health record refers to a structured data set of individualized health benchmark information for the elderly that is pre-stored during the user initialization phase through methods such as importing records from medical institutions, filling in information by family members, or migrating historical health data. This includes the average systolic blood pressure over the past 30 days, the average diastolic blood pressure over the past 30 days, the median resting heart rate, the 5th percentile of blood oxygen saturation, the circadian rhythm fitting curve of body temperature, and the historical average fasting blood glucose, which are used as personal health baseline indicators.

[0042] S3.2: The three-level warning criteria of green, yellow and red are used to comprehensively judge the degree of abnormality of physiological parameters, changes in behavioral patterns and medication adherence in the health status deviation data, and generate health risk assessment results; Specifically, the system extracts the number of abnormal physiological parameters, the duration of each abnormal physiological parameter, behavioral pattern change markers, and medication adherence assessment results from health status deviation data. It then determines whether the number of abnormal physiological parameter types is greater than or equal to three; if so, it proceeds to the red alert assessment process. If the number of abnormal physiological parameter types is less than three but greater than or equal to one, and at least one abnormal physiological parameter has a duration of greater than or equal to three days, it proceeds to the yellow alert assessment process. If the number of abnormal physiological parameter types is zero, or if abnormalities exist but all abnormalities have a duration of less than three days, it proceeds to the green alert assessment process. In the red alert assessment process, a red health risk assessment result is generated based on behavioral pattern change markers such as reduced activity or sleep disturbances, and a medication adherence assessment result below a preset standard. In the yellow alert assessment process, a yellow health risk is generated if behavioral pattern change markers include changes in daily activity patterns or occasional missed medication doses. In the green alert assessment process, a green health risk is generated if medication adherence is normal and there are no behavioral pattern changes. Finally, a red, yellow, or green health risk is output as the health risk assessment result.

[0043] It should be noted that the green-yellow-red three-level warning standard is a classification rule set based on the number of abnormal physiological parameters in the health status deviation data, the duration of each abnormal physiological parameter, the markers of behavioral pattern changes, and the medication adherence assessment results: Green warning corresponds to the situation where there are no abnormal physiological parameters or the abnormality lasts for less than three days, there are no behavioral pattern changes, and the medication adherence assessment result is normal; Yellow warning corresponds to the situation where there are one or two abnormal physiological parameters and at least one abnormality lasts for more than three days, or is accompanied by changes in daily activity patterns and the medication adherence assessment result is occasional missed doses; Red warning corresponds to the situation where there are more than three types of abnormal physiological parameters, or there are markers of behavioral pattern changes such as reduced activity level and sleep disorders, and the medication adherence assessment result is lower than the preset standard.

[0044] The preset standard is based on the required frequency and dosage of medication as specified in the Clinical Guidelines for the Management of Chronic Diseases in the Elderly and the Individualized Medication Plan. The range of values ​​is as follows: when the ratio of the actual number of times medication is taken to the number of times medication should be taken is less than 0.8, the medication adherence judgment result is considered to be below the preset standard.

[0045] S3.3: Perform keyword matching and sentiment dictionary analysis on rich emotional interaction data and behavioral data, extract sentiment keywords, and identify the user's current sentiment theme; Specifically, rich emotional interaction data and behavioral data are extracted from a structured multimodal time-series dataset; the rich emotional interaction data is segmented to obtain word sequences; each word in the word sequence is matched with an emotional dictionary, and words that match the emotional dictionary are selected as candidate emotional words; part-of-speech tagging and semantic role analysis are performed on the candidate emotional words to exclude emotional words in negative contexts and retain effective emotional words; the effective emotional words are correlated with the behavior type, behavior frequency, and behavior time in the behavioral data to determine whether there are emotional reinforcement behavior patterns (e.g., prolonged sitting accompanied by sadness-related emotional words, frequent looking at old photos accompanied by nostalgia-related emotional words); based on the effective emotional words and associated behavioral patterns, the core emotional tendency currently expressed by the user is summarized; the core emotional tendency is mapped to predefined emotional theme categories, including nostalgia, loneliness, joy, anxiety, or need for social support, to identify the user's current emotional theme.

[0046] It should be noted that the emotional theme categories are defined based on the common psychological states and life scenarios of the elderly, combined with psychological theories and the needs of elderly care practice. Specifically, they include five emotional themes: nostalgia, loneliness, joy, anxiety, and the need for social support. The semantic boundaries of each emotional theme are defined through an expert-annotated corpus and are solidified into a set of emotional theme categories during the system initialization phase.

[0047] Emotional keywords refer to effective words retained after segmentation, matching with an emotional dictionary, and filtering through negative contexts from emotionally rich interactive data. Specifically, they include the following three categories: First, basic emotional expression vocabulary, such as happy, sad, afraid, and angry; second, emotional expressions specific to the lives of the elderly in their daily lives, such as homesickness, loneliness, no one to accompany them, feeling useless in old age, reminiscing about the past, and missing their spouse; third, keywords directly corresponding to five predefined emotional themes, including nostalgia (such as "the past," "those days," old photos, joining the army, getting married, retiring, hometown, and old classmates), and loneliness (such as being alone, having no one to talk to, feeling cold and lonely, and needing social interaction because children are not around). Supportive categories include wanting to see children, make phone calls, meet up with old friends, and have meals together; anxiety categories include insomnia, heart palpitations, taking the wrong medication, and high blood pressure; and joyful categories include grandchildren visiting, being happy today, singing opera, and receiving phone calls. Emotional keywords must simultaneously meet two conditions: First, they must appear in word sequences rich in emotional interaction data from voice dialogue transcription, handwriting input, or touchscreen interaction logs; second, they must be confirmed to be in a non-negative context through part-of-speech tagging and semantic role analysis (e.g., excluding happiness within unhappiness). Ultimately, they are used as valid emotional words to be associated with behavioral data (e.g., sitting still duration, frequency of photo viewing) to support the identification of the user's current emotional theme.

[0048] S3.4: Associate the user's current emotional theme with important life events in the experience graph, and determine whether the user is currently in an emotional state of nostalgia, loneliness, or need for social support, and generate emotional need recognition results.

[0049] Specifically, semantic matching is performed between the user's current emotional theme and important life milestones in their experience graph to determine whether the user's current emotional theme is nostalgia, loneliness, or a need for social support. If the user's current emotional theme is nostalgia, important life milestones in the experience graph containing keywords such as relatives, hometown, wedding, military service, and retirement are retrieved, and it is confirmed whether there are any upcoming anniversaries or seasonal triggers. If the user's current emotional theme is loneliness, important life milestones in the experience graph during periods of sparse social relationships are retrieved, and this is verified by combining records of no social interaction for three consecutive days in behavioral data. If the user's current emotional theme is a need for social support... The system retrieves important life events in the experience graph that were previously marked as having high emotional value and were associated with family members or old friends, and checks whether any related individuals have not appeared in the recent interaction records. Based on the above matching and verification results, it determines whether the user is currently in an emotional state of nostalgia, loneliness, or need for social support. A structured triple is generated, which includes the emotional state category (nostalgia, loneliness, or need for social support), a list of associated event identifiers (a set of IDs of important life events that were successfully matched in the experience graph), and the emotional need identification results based on trigger criteria (including an upcoming anniversary, the number of consecutive days without social interaction, and the absence of recent contextual verification information of related individuals).

[0050] S4: Based on the results of emotional needs recognition, retrieve related memory materials from the individual's full-dimensional knowledge base, and generate multimodal care content through AIGC; S4.1: Semantically match the emotion tags in the emotion need identification results with the metadata tags of multimodal memory materials in the personal full-dimensional knowledge base, and retrieve candidate multimodal memory materials associated with event nodes in the experience graph from the memory association network; Specifically, emotional tags are extracted from the emotional needs identification results; based on the event nodes in the experience graph, multimodal memory materials associated with each event node and their corresponding metadata tags are obtained through a memory association network; the semantic similarity between the emotional tags and the emotional type in the metadata tags of each multimodal memory material is calculated; multimodal memory materials with semantic similarity greater than the semantic similarity threshold are selected to form a candidate multimodal memory material set.

[0051] The semantic similarity calculation expression is: ; Candidate multimodal memory material set filtering expression: ; in, Indicates the sentiment label and the first Semantic similarity scores between sentiment types of multimodal memory materials; This represents the emotional tags extracted from the emotional needs identification results; express The sentiment type in the corresponding metadata tag; Indicates the semantic similarity threshold; This represents the set of candidate multimodal memory materials; Indicates the first Multimodal memory materials.

[0052] It should be noted that the semantic similarity threshold is preset by offline optimization of the labeled sample set during the initialization stage, based on the characteristics of emotional expression of the elderly and the requirements for recall accuracy of memory materials. The value range is from 0.6 to 0.85.

[0053] S4.2: Calculate the semantic similarity between the sentiment tags and the metadata tags of the candidate multimodal memory materials, sort them according to the similarity scores, and select the multimodal memory material with the highest score as the selected multimodal memory material; Specifically, the process involves reading multimodal memory materials and their corresponding metadata tags one by one from the candidate multimodal memory material set; extracting the sentiment type field from each metadata tag; inputting the sentiment tags from the sentiment demand recognition results and each sentiment type field into the semantic similarity calculation model to obtain the corresponding semantic similarity score; pairing all semantic similarity scores with the corresponding multimodal memory materials to form a score-material pair list; sorting the score-material pair list from high to low according to the semantic similarity score; and selecting the multimodal memory material corresponding to the first score-material pair in the sorted list as the selected multimodal memory material.

[0054] It should be noted that the data collection of emotional expression pairs among the elderly includes positive sample pairs of emotional labels and corresponding event descriptions, as well as negative sample pairs of irrelevant labels and descriptions, forming an emotional corpus dataset for the elderly. Each text pair is segmented and vectorized, and sentence vectors are generated using pre-trained Chinese word vectors (such as Chinese-BERT-wwm). A Siamese network structure is constructed, and the two sentence vectors are input into a weighted encoder to output semantic similarity scores. End-to-end training is performed using a contrastive loss function or a triplet loss function, so that the similarity scores of positive sample pairs approach 1 and the similarity scores of negative sample pairs approach 0. Model performance is monitored on a validation set, and training stops when the F1 score reaches a convergent and stable state. The trained model is then solidified into a semantic similarity calculation model.

[0055] S4.3: Use AIGC to generate multimodal companionship content by using selected multimodal memory materials as contextual cues.

[0056] Specifically, the text descriptions, image metadata, and related event information from the selected multimodal memory materials are structurally concatenated to form contextual cues; these contextual cues are then input into AIGC; AIGC generates multimodal companionship content containing voice narration, image-text summaries, and emotional statements based on the contextual cues; the generated multimodal companionship content undergoes format standardization processing, converting the voice narration into a WAV format audio file with a sampling rate of 16kHz, the image-text summary into a JPEG format image file, and the emotional statements into a UTF-8 encoded text file; finally, the multimodal companionship content is output.

[0057] S5: Based on health risk assessment results and multimodal care content, construct structured interactive scripts and collaboratively execute health diagnosis assistance and care content through cross-device scheduling.

[0058] S5.1: Based on the warning level in the health risk assessment results, generate corresponding health prompts, set the upper limit of the playback time of the care content, arrange the health prompts to be played before the start of the multimodal care content, and generate health intervention scheduling instructions; Specifically, the warning level is read from the health risk assessment results; if the warning level is green, a daily health reminder is generated as a health prompt; if the warning level is yellow, a health attention suggestion is generated as a health prompt; if the warning level is red, an emergency medical treatment prompt is generated as a health prompt; the playback time limit for the companionship content is set according to the warning level: 15 minutes for a green warning level, 10 minutes for a yellow warning level, and 5 minutes for a red warning level; the health prompt is played before the multimodal companionship content begins, forming a health intervention programming instruction that includes health prompts, multimodal companionship content, and playback sequence constraints.

[0059] S5.2: Integrate health tips and multimodal companionship content according to their playback order and duration limit to obtain a structured interactive script; Specifically, the system extracts health prompts, multimodal companionship content, and the maximum playback time of the companionship content from the health intervention scheduling instructions; it sets the start time of the health prompts to midnight and calculates the end time of the health prompts; it sets the start time of the multimodal companionship content to the end time of the health prompts, and sets the end time of the multimodal companionship content to the start time of the multimodal companionship content plus the maximum playback time of the companionship content; it then combines the health prompts, multimodal companionship content, and their respective start and end times in chronological order to generate a structured interactive script containing content identifiers, content types, start times, and end times.

[0060] S5.3: Parse device instructions in the structured interactive script, distribute different content to the corresponding target terminal devices, and output the content-device mapping and distribution results; Specifically, the system reads content identifiers, content types, playback start times, and playback end times from the structured interactive script; it determines the content category based on the content type: if the content type is voice narration, it assigns the content corresponding to the content identifier to the voice broadcasting device; if the content type is a text and image summary, it assigns the content corresponding to the content identifier to the display device; if the content type is emotional statements, it assigns the content corresponding to the content identifier to a smart speaker or companion robot; it pairs each piece of content with the assigned target terminal device to form a content-device mapping relationship; it summarizes all content-device mapping relationships and outputs the content-device mapping distribution results.

[0061] S5.4: Based on the mapping and distribution results between content and devices, through cross-device scheduling, each target terminal device starts playback according to the interaction sequence in the structured interactive script, and maintains playback synchronization and timing alignment.

[0062] Specifically, the system reads the target terminal device corresponding to each piece of content and its playback start time in the structured interactive script from the content-device mapping and distribution results; generates a device playback instruction for each target terminal device, containing content data, playback start time, and playback end time; sends all device playback instructions to the corresponding target terminal devices; after receiving the device playback instruction, each target terminal device aligns its local clock according to the playback start time and waits for the trigger moment under a unified time base; when the global time reaches zero, each target terminal device synchronously starts timing; when the local time reaches its respective playback start time, each target terminal device starts playing the corresponding content; during playback, each target terminal device continuously compares its local playback progress with the timeline in the structured interactive script; if a playback offset exceeds a preset tolerance threshold, playback is paused, and the playback progress is realigned based on the unified time base; after all content playback is completed, each target terminal device stops outputting, realizing the collaborative execution of health diagnosis assistance and companionship content.

[0063] It should be noted that the tolerance threshold is preset by offline calibration of the offset data played on multiple devices during the system initialization phase, based on the clock drift characteristics of the target terminal device and the cross-device synchronization accuracy requirements. The value range is ±200 milliseconds to ±500 milliseconds.

[0064] This embodiment also provides an AIGC-based health diagnosis and care system for the elderly, including: The data preprocessing module is used to collect multi-source heterogeneous data of the elderly in real time through smart sensors, perform preprocessing, and generate structured multimodal time series datasets. The knowledge fusion module is used to integrate structured multimodal time-series datasets, pre-stored life experience data, and health records to build and dynamically update a personal full-dimensional knowledge base; The state recognition module is used to analyze health risk status and emotional need status based on structured multimodal time series datasets and personal full-dimensional knowledge base, generate health risk assessment results, and identify emotional needs. The content generation module is used to retrieve related memory materials from the personal full-dimensional knowledge base based on the emotional needs recognition results, and generate multimodal companionship content through AIGC; The collaborative execution module is used to construct structured interactive scripts based on health risk assessment results and multimodal care content, and to collaboratively execute health diagnosis assistance and care content through cross-device scheduling.

[0065] In summary, this invention achieves unified modeling of experience graphs and memory association networks by constructing and dynamically updating a personal full-dimensional knowledge base, supporting accurate reasoning of pattern recognition within the context of an individual's life course; and by constructing structured interactive scripts, it enables cross-modal synchronous execution of health interventions and emotional companionship at the content, temporal, and device levels, allowing pattern recognition not only to be used for state discrimination but also to drive proactive care with memory continuity and mind-body synergy, ultimately achieving a personalized elderly care closed loop that deeply integrates with the individual's background.

[0066] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for health diagnosis and care for the elderly based on AIGC, characterized in that: include, The system collects multi-source heterogeneous data from elderly people in real time using smart sensors, performs preprocessing, and generates a structured multimodal time-series dataset. By integrating structured multimodal time-series datasets, pre-stored life experience data, and health records, a personal full-dimensional knowledge base is constructed and dynamically updated. Based on a structured multimodal time-series dataset and a personal full-dimensional knowledge base, we analyze health risk status and emotional needs status, generate health risk assessment results, and identify emotional needs. Based on the results of emotional needs identification, relevant memory materials are retrieved from the individual's full-dimensional knowledge base, and multimodal care content is generated through AIGC. Based on health risk assessment results and multimodal care content, a structured interactive script is constructed, and health diagnosis assistance and care content are executed collaboratively through cross-device scheduling.

2. The method for health diagnosis and care of the elderly based on AIGC as described in claim 1, characterized in that: The multi-source heterogeneous data includes physiological data, behavioral data, environmental data, and emotionally rich interactive data; The preprocessing includes data cleaning, noise reduction, calibration, and timing alignment.

3. The method for health diagnosis and care of the elderly based on AIGC as described in claim 2, characterized in that: The steps for integrating structured multimodal time-series datasets, pre-stored life experience data, and health records are as follows: The structured multimodal time-series dataset is time-aligned and semantically matched with pre-stored life experience data to obtain the life experience fusion result; The structured multimodal time-series dataset is compared and analyzed with pre-stored health records, and the difference between the current value of physiological parameters and the individual's health baseline index is calculated to obtain the health record fusion results.

4. The method for health diagnosis and care of the elderly based on AIGC as described in claim 3, characterized in that: The steps for constructing and dynamically updating a personal full-dimensional knowledge base are as follows. Based on the integration results of life experiences and health records, an experience map and memory association network are constructed to generate an initial personal full-dimensional knowledge base; Based on a structured multimodal time-series dataset, the initial personal full-dimensional knowledge base is dynamically updated to generate a dynamically updated personal full-dimensional knowledge base.

5. The method for health diagnosis and care of the elderly based on AIGC as described in claim 4, characterized in that: The steps for generating the health risk assessment results are as follows: Based on structured multimodal time-series datasets and pre-stored health records, the deviation of the current health status from the individual's health baseline indicators is calculated to obtain health status deviation data. The three-tiered early warning system (green-yellow-red) is used to comprehensively assess the degree of abnormality in physiological parameters, changes in behavioral patterns, and medication adherence in health status deviation data, thereby generating health risk assessment results.

6. The method for health diagnosis and care of the elderly based on AIGC as described in claim 5, characterized in that: The steps for identifying emotional needs are as follows. We will perform keyword matching and sentiment dictionary analysis on emotionally rich interactive data and behavioral data, extract sentiment keywords, and identify the user's current sentiment theme. The system associates a user’s current emotional theme with important life events in their experience graph, and determines whether the user is currently in an emotional state of nostalgia, loneliness, or need for social support, generating an emotional need identification result.

7. The method for health diagnosis and care of the elderly based on AIGC as described in claim 6, characterized in that: The steps for generating multimodal care content are as follows: The emotional tags in the emotional needs identification results are semantically matched with the metadata tags of multimodal memory materials in the personal full-dimensional knowledge base, and candidate multimodal memory materials associated with event nodes in the experience graph are retrieved from the memory association network. Calculate the semantic similarity between sentiment tags and metadata tags of candidate multimodal memory materials, sort them according to similarity scores, and select the multimodal memory material with the highest score as the selected multimodal memory material; AIGC uses selected multimodal memory materials as contextual cues to generate multimodal companionship content.

8. The method for health diagnosis and care of the elderly based on AIGC as described in claim 7, characterized in that: Based on health risk assessment results and multimodal care content, a structured interactive script is constructed, and the steps are as follows: Based on the warning level in the health risk assessment results, corresponding health prompts are generated, and the maximum playback time of the care content is set. The health prompts are played before the multimodal care content begins, and health intervention scheduling instructions are generated. The health tips and multimodal companionship content are sequentially integrated according to their playback order and duration limit to obtain a structured interactive script.

9. The method for health diagnosis and care of the elderly based on AIGC as described in claim 8, characterized in that: The steps for collaboratively executing health diagnosis assistance and care services through cross-device scheduling are as follows: Parse device instructions in structured interactive scripts, distribute different content to corresponding target terminal devices, and output the content-device mapping and distribution results; Based on the mapping and distribution results between content and devices, cross-device scheduling enables each target terminal device to start playback according to the interaction sequence in the structured interactive script, and maintain playback synchronization and timing alignment.

10. An AIGC-based health diagnosis and care system for the elderly, based on the AIGC-based health diagnosis and care method for the elderly as described in any one of claims 1 to 9, characterized in that: include, The data preprocessing module is used to collect multi-source heterogeneous data of the elderly in real time through smart sensors, perform preprocessing, and generate structured multimodal time series datasets. The knowledge fusion module is used to integrate structured multimodal time-series datasets, pre-stored life experience data, and health records to build and dynamically update a personal full-dimensional knowledge base; The state recognition module is used to analyze health risk status and emotional need status based on structured multimodal time series datasets and personal full-dimensional knowledge base, generate health risk assessment results, and identify emotional needs. The content generation module is used to retrieve related memory materials from the personal full-dimensional knowledge base based on the emotional needs recognition results, and generate multimodal companionship content through AIGC; The collaborative execution module is used to construct structured interactive scripts based on health risk assessment results and multimodal care content, and to collaboratively execute health diagnosis assistance and care content through cross-device scheduling.

Images