Personalized health auxiliary management system based on artificial intelligence

Abstract

The invention relates to the technical field of old people health management, and discloses a personalized health auxiliary management system based on artificial intelligence, which comprises a multi-mode sensing module, an intelligent preprocessing module, a dynamic knowledge graph module, a personalized inference engine module and a self-adaptive communication and interaction module. The multi-mode sensing module is used for acquiring physiological, environmental and behavioral data; the intelligent preprocessing module desensitizes and cleans data through federal learning, and extracts core features; the dynamic knowledge graph module constructs and updates an exclusive graph containing basic and dynamic nodes and associated edges; the personalized inference engine module generates a layered intervention scheme; and the adaptive interaction module adapts to the cognitive ability and synchronizes the scheme. According to the invention, health information integration and dynamic intervention are realized, the management accuracy and compliance are improved, and the method is suitable for home non-clinical scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of health management for the elderly, and in particular to an individualized health auxiliary management system based on artificial intelligence. BACKGROUND

[0002] Health management for the elderly has gradually become a focus of society, and the existing health management methods for the elderly have many limitations: health data collection is scattered and not integrated, physiological parameters, daily behaviors, etc. Data is recorded independently by a single device, which is difficult to form a correlation analysis, making it difficult for medical staff to fully assess the health status of the elderly. There is a contradiction between data processing and privacy protection. Although centralized data management is convenient for analysis, it has the risk of sensitive health information leakage, and distributed management also makes it difficult to achieve collaborative optimization of multi-source data. Health intervention programs lack individualization and dynamic adaptation. Traditional management mostly uses standardized recommendations, ignoring individual differences such as medical history and cognitive ability of the elderly, and cannot adjust intervention strategies according to real-time health status, resulting in low intervention compliance. The interaction method does not match the needs of the elderly. Existing systems mostly use unified operation interfaces without considering the cognitive decline of the elderly, such as memory loss and decreased vision, increasing the use of obstacles. The risk response is lagging behind. The identification of potential health risks relies on manual judgment, making it difficult to achieve efficient response from prevention to emergency. SUMMARY

[0003] The present application aims to provide an individualized health auxiliary management system based on artificial intelligence to achieve individualized health management for the elderly through a closed-loop architecture of perception, processing, modeling, reasoning, and interaction.

[0004] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: The individualized health auxiliary management system based on artificial intelligence comprises: A multi-modal perception module for collecting physiological parameters, environmental information, and behavior capture information of the elderly; An intelligent preprocessing module for desensitizing and cleaning the raw data collected by the multi-modal perception module through a federated learning framework, filtering motion artifact interference using a time series network with an attention mechanism, and extracting physiological parameter features, environmental information features, and behavior features; A dynamic knowledge graph module for constructing an exclusive health knowledge graph for the elderly based on the extracted features, containing basic nodes, dynamic nodes, and associated edges, and updating node attributes and association strength through incremental learning; An individualized reasoning engine module for constructing a health management strategy generation model based on deep reinforcement learning, with the real-time state vector of the knowledge graph as input and a hierarchical intervention scheme as output; the hierarchical intervention scheme includes primary prevention recommendations, secondary early warning interventions, and tertiary emergency coordination; An adaptive communication and interaction module is configured to dynamically adjust the interaction mode according to the cognitive ability evaluation result of the elderly, provide a real-time communication channel for the elderly and family members or medical staff, and synchronize the hierarchical intervention scheme to the family members or medical staff.

[0005] The principle and advantages of the scheme are as follows: in actual application, the physiological parameters, environmental information and behavior capture information of the elderly are comprehensively collected through the multi-modal perception module to provide a data basis for subsequent processing; the multi-source data is desensitized and cleaned by federated learning, the interference is filtered by the time sequence network combined with the attention mechanism, and the core features such as physiological rhythm, behavior cycle and environmental correlation are extracted; the knowledge graph containing basic information, dynamic health status and correlation is constructed based on the extracted features, and the node attributes and correlation strength are updated in real time through incremental learning to realize dynamic integration of health information; a deep reinforcement learning model is used to take the real-time state of the knowledge graph as input to generate a hierarchical intervention scheme adapted to the health status of the elderly; the interaction mode is dynamically adjusted according to the cognitive ability of the elderly, and the intervention scheme is synchronized to the family members or medical staff to ensure the accessibility and synergy of the intervention. Through multi-modal perception and dynamic knowledge graph, the application breaks the isolated state of physiological, behavioral and environmental data, realizes three-dimensional modeling of health information, and makes the intervention scheme respond to the changes in health status in real time and adapt to individual differences of the elderly through the reasoning engine based on deep reinforcement learning and the knowledge graph based on incremental learning. The adaptive interaction module matches the cognitive ability of the elderly, reduces the operation obstacles, and improves the execution degree of the intervention scheme through the cooperation of family members and medical staff. The hierarchical intervention mechanism combines the correlation relationship of the knowledge graph to realize the whole-chain management from prevention to emergency and reduce the risk of home health. The application effectively solves the problems of "data dispersion, intervention rigidity and interaction inadaptation" of the existing system in the health management of the elderly, and is especially suitable for intelligent health assistance in non-clinical home scenarios.

[0006] Preferably, as an improvement, the basic node includes static information, and the static information includes age, gender and medical history; The dynamic node includes the extracted features, including the extracted physiological parameter features, environmental information features and behavior features. The correlation edge includes the correlation strength between the features.

[0007] Technical effect: by combining the static and dynamic nodes, the foundation of personalized management is laid while reflecting the changes in health status in real time, the graph can capture the subtle changes in health status through real-time updating of the features, and the correlation edge is adjusted to realize the leap from isolated features to overall health assessment.

[0008] Preferably, as an improvement, the dynamic knowledge graph module includes: A feature mapping sub-module is configured to convert the extracted features into entities and attributes of the graph. an updating submodule configured to determine an updating period and update dynamic nodes and associated edges of the graph based on newly extracted features.

[0009] Technical effects: Facilitate updating the knowledge graph structure according to the condition of the elderly.

[0010] Preferably, as an improvement, the incremental learning includes: an entity alignment submodule configured to automatically associate the same type of health events by calculating semantic similarity between newly collected data and existing nodes through a graph attention network; a relationship reasoning submodule configured to update conditional probability of the associated edges based on a Bayesian network; a decay mechanism submodule configured to perform weight decay on historical data exceeding a preset time window.

[0011] Technical effects: Facilitate keeping the knowledge graph always up-to-date and accurate, and provide reliable dynamic basis for health management decisions.

[0012] Preferably, as an improvement, the incremental learning further includes an associated node attribute updating submodule configured to automatically update attributes of associated nodes based on real-time updated node attributes.

[0013] Technical effects: Facilitate further improving the completeness and accuracy of the knowledge graph updating.

[0014] Preferably, as an improvement, the deep reinforcement learning adopts a double-delay deep deterministic policy gradient algorithm; the state space contains a plurality of feature vectors, the action space contains a combination strategy of basic intervention schemes, and the reward function is a composite index.

[0015] Technical effects: Facilitate upgrading from single-index management to comprehensive benefit optimization; mixed training of simulated environment and real data improves the robustness of the model, ensuring that the optimal intervention scheme can still be generated in complex home scenarios, effectively balancing the health management effect and the convenience of the elderly.

[0016] Preferably, as an improvement, the cognitive ability assessment includes memory, executive function, language ability, visual spatial ability, and calculation ability, and the interactive level is automatically matched according to the assessment results.

[0017] Technical effects: Facilitate comprehensive assessment of the cognitive ability of the elderly.

[0018] Preferably, as an improvement, the dynamic knowledge graph module further includes a warning triggering submodule configured to combine the weight of the associated edges between nodes and the node attribute combination features as a warning triggering condition, specifically including: preset multi-level warning thresholds, setting a basic weight threshold for different types of associated edges and setting a dynamic threshold for node attribute combination features; Real-time monitoring of the change of the correlation edge weight, triggering an early warning when the correlation edge weight exceeds the basic threshold, and pushing an intervention scheme to the medical staff front-end module through the adaptive communication and interaction module; Recognizing the abnormal pattern of node attribute combination, automatically activating the three-level emergency coordination when the combination feature exceeds the threshold, obtaining the real-time location, issuing a prompt to the elderly through the adaptive communication and interaction module, and sending the risk level, the associated node feature screenshot and the positioning information to the preset emergency contact.

[0019] Technical effect: It is convenient to link knowledge graph and early warning function, by taking the node correlation relationship and attribute combination feature as the early warning trigger condition, combining the multi-level threshold system, the precise grading early warning of health risk is realized, which not only ensures that potential problems are discovered and professional adjustment schemes are pushed in time, but also quickly links the emergency response to high-risk combination, improves the pertinence and timeliness of early warning, and enhances the safety of health management of the elderly. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The structure diagram of the individualized health auxiliary management system based on artificial intelligence. DETAILED DESCRIPTION

[0021] The following will be further described in detail through specific embodiments: The embodiments are basically as shown in the accompanying drawings: Figure 1 The individualized health auxiliary management system based on artificial intelligence comprises: ​A multi-modal perception module is used to collect physiological parameters, environmental information and behavior capture information of the elderly. Specifically, the wearable physiological sensor array, the environmental sensor group and the behavior capture device are used for collection. The wearable physiological sensor array integrates flexible electrode sheet and optical sensing module to collect blood pressure, heart rate and blood glucose data of the elderly in real time. The flexible electrode sheet of the wearable physiological sensor array adopts graphene-silver nanowire composite material with a thickness of ≤0.1 mm and a deformation adaptability of ≥20%. The optical sensing module integrates 660 nm / 940 nm dual-wavelength LED and avalanche photodiode to realize synchronous monitoring of blood oxygen saturation and non-invasive blood glucose, and the sampling frequency is dynamically adjusted according to the activity intensity, such as 1 Hz in resting state and 10 Hz in motion state. The environmental sensor group is deployed in the key areas of the living space to monitor temperature and humidity, light intensity and air pollutant concentration. The behavior capture device adopts the fusion technology of millimeter wave radar and computer vision to identify daily activity types (including eating, taking medicine, gait and sitting behavior) and abnormal motion patterns (including falling risk posture and activity reduction). The millimeter wave radar of the behavior capture device works in the 77 GHz frequency band to extract human micro-motion features through micro-Doppler effect, and the spatiotemporal alignment is performed with the skeleton key point detection result of computer vision. The behavior recognition calculation is completed in the local terminal, and only the feature vector of the recognition result is uploaded to the cloud.

[0022] An intelligent preprocessing module is used to desensitize and clean the original data collected by the multi-modal perception module through the federated learning framework, to filter motion artifact interference by using a time sequence network with attention mechanism, and to extract physiological parameter features, environmental information features and behavior features. Specifically, each terminal (such as a wearable device and a home sensor) first desensitizes the collected original data locally, for example, differentially private disturbance is performed on specific numerical values in physiological parameters (such as a random offset of ±2 mmHg for blood pressure values), and time stamps in behavior data are fuzzed (accurate to the hour level) to avoid directly exposing real-time activity details of the elderly. The local terminal only uploads the feature parameters after data cleaning to the cloud, the cloud updates the global model by aggregating parameters of multiple terminals, and then the optimized model is distributed to each terminal. Motion artifacts (such as abnormal heart rate data caused by shaking of the wearable device) are common interference in physiological monitoring. The attention mechanism dynamically allocates weights to focus on effective signals, converts continuous physiological data (such as heart rate curve within 10 minutes) into time sequence, captures time dependence through LSTM network, and introduces attention layer to allocate lower weight to periods of intense motion (such as stepping action in gait monitoring) and higher weight to resting periods, thereby reducing the influence of interference signals. The motion sensor data of the wearable device, such as acceleration, is used as auxiliary input, and when the acceleration value exceeds the threshold (such as >0.5g), the filtering strength of the same period physiological data is automatically enhanced to further eliminate motion artifacts.

[0023] The physiological parameter feature includes a circadian rhythm feature of the physiological parameter. The continuous physiological data (such as blood pressure, blood glucose) is segmented by using a fixed step (such as 24 hours) sliding window, the mean value, peak value, valley value and fluctuation amplitude in the window are calculated, and the circadian rhythm mode (such as morning blood pressure peak, night blood glucose valley) is identified. The time domain data is converted to the frequency domain by Fourier transform, and the energy proportion of the periodic component is detected. If the proportion is greater than 60%, it is determined that there is a significant circadian rhythm. The circadian rhythm feature of the physiological parameter can reflect the basic health status and disease control of the elderly. For example, by monitoring the circadian blood pressure rhythm, the abnormal mode (such as non-dipper hypertension) of hypertension can be identified early, and the basis for personalized medication time adjustment is provided.

[0024] The behavior feature includes periodic feature and non-periodic feature. The periodic feature extraction includes the regularity of daily activity data (such as eating, sleeping, activity and other behaviors), such as daily eating frequency, sleep time, activity amount, and medication time. The periodic mode (such as fixed meal time, regular sleep) is identified by K-means algorithm. The deviation (such as eating interval more than mean ± 2 hours) of the current behavior feature from the historical periodic feature is calculated. If the deviation exceeds the threshold, it is marked as abnormal. Through the regularity feature of daily activity data, the self-care ability and health habit of the elderly can be evaluated. When the behavior period deviates significantly, potential health risks can be prompted, such as a 20% decrease in sleep duration for three consecutive days, indicating a potential depression tendency or prodromal symptoms. The non-periodic feature includes abnormal motion, such as fall risk posture.

[0025] The environmental information feature includes environmental parameters and associated features. The correlation between environmental parameters and physiological / behavioral data is calculated by Pearson coefficient. For example, potential correlation features such as “humidity > 60% increases fall risk” and “light duration < 4 hours is related to sleep quality decline” are identified. By identifying the potential impact of the environment on health, suggestions for home environment optimization can be provided.

[0026] The dynamic knowledge graph module is used to construct a health knowledge graph for the elderly based on the extracted features. The health knowledge graph includes basic nodes, dynamic nodes and associated edges, and updates node attributes and association strength through incremental learning.

[0027] The basic node includes static information, and the static information includes age, gender, and medical history. The static information is used as the skeleton of the knowledge graph. The dynamic node includes the extracted physiological parameter feature, environmental information feature and behavior feature. The associated edge includes the association strength between the features.

[0028] The dynamic knowledge graph module includes: The feature mapping submodule is configured to convert the extracted features into entities and attributes of the graph, such as "sleep cycle features" being mapped to attributes of a "sleep node" (e.g., "deep sleep proportion = 25%" and "sleeping time fluctuation = ± 30 minutes"); and "environmental correlation features" being mapped to correlation edges between "environment nodes" and "physiological nodes" (e.g., "humidity > 60% corresponding to a 15% increase in the risk of high blood pressure").

[0029] The updating submodule is configured to determine an updating period, which is preferably 24 hours in this embodiment, and update the dynamic nodes and correlation edges of the graph based on newly extracted features, including adding, deleting, and modifying operations.

[0030] The incremental learning includes an entity alignment submodule, a relationship reasoning submodule, and a decay mechanism submodule.

[0031] The entity alignment submodule is configured to calculate the semantic similarity between newly collected data and existing nodes by using a graph attention network, and automatically associate the same type of health events. Specifically, first, a feature vector is constructed to convert the newly collected data (e.g., "2025-07-10 blood glucose 8.5 mmol / L") and the existing nodes (e.g., "2025-07-01 blood glucose 8.3 mmol / L") into 128-dimensional feature vectors, including numerical values, timestamps, and collection scene attributes. Then, the attention weight is calculated, and the correlation weight between the two nodes is calculated by the attention layer of GAT:

[0032] wherein, is the feature vector of the new data, is the feature vector of the existing node, and a is a learnable parameter, is the attention weight of node i and j (range 0-1). When is greater than a preset threshold (e.g., 0.7), it is determined as the same type of health event (e.g., "postprandial hyperglycemia"), and the new data is automatically associated with the corresponding node, and the node attribute is updated (e.g., "high blood glucose frequency in the past 7 days + 1").

[0033] The relationship reasoning submodule is configured to update the conditional probability of the correlation edges based on a Bayesian network. Specifically, the Bayesian network structure is initialized to construct a directed acyclic graph (DAG) of "behavior-health indicators", such as "high-salt diet indicating high blood pressure" and "insufficient exercise indicating blood glucose fluctuation", and the initial conditional probability P(B|A); the conditional probability is updated by using the Bayesian estimation formula:

[0034] wherein, is the number of samples in which A occurs and B occurs, is the total number of samples in which A occurs, and is a smoothing coefficient, used to avoid zero probability, usually set to 1. For example, among the newly added 10 pieces of data, "high-salt diet" appears 6 times, of which 5 times is accompanied by "blood pressure rise", then is 0.75. The conditional probability P(B|A) is linearly mapped to the association edge weight W:

[0035] is the maximum probability of the same association.

[0036] The decay mechanism submodule performs weight decay on historical data that exceeds the preset time window, preferentially retains recent key information, and ensures that the knowledge graph preferentially reflects the recent health status, avoiding interference from historical redundant information. Specifically, the effective time window is determined, and for data that exceeds the effective time window, an exponential decay function is used to calculate the weight:

[0037] where t is the current time, T is the data collection time, and λ is the decay coefficient (the shorter the window length, the larger λ, such as 30-day window λ=0.05, 5-year window λ=0.0004).

[0038] Set a decay lower limit for key data to ensure that core health information is not completely weakened, such as setting a decay lower limit (e.g., W≥0.5) for key data such as myocardial infarction history and drug allergies.

[0039] The incremental learning further includes an associated node attribute updating submodule for automatically updating the attributes of the associated nodes based on the real-time updated node attributes, so as to further improve the completeness and accuracy of the knowledge graph update.

[0040] The personalized reasoning engine module is used to construct a health management strategy generation model based on deep reinforcement learning, with the input being a real-time state vector of the knowledge graph and the output being a hierarchical intervention scheme; the hierarchical intervention scheme includes first-level prevention suggestions, second-level warning interventions, and third-level emergency coordination, the first-level prevention suggestions include diet adjustment and exercise intensity adaptation schemes for health indicator fluctuation trends, the second-level warning interventions include medication time optimization and medical resource reservation prompts for potential risks, and the third-level emergency coordination includes response priority sorting of linked emergency contacts and community medical stations. The deep reinforcement learning uses a double-delay deep deterministic policy gradient algorithm, the state space contains 128-dimensional feature vectors, specifically, 64-dimensional physiological features, 32-dimensional behavior features, 16-dimensional environmental features, and 16-dimensional historical intervention effect features, the action space contains 256 combinations of basic intervention schemes, and the reward function is a composite index including health indicator compliance rate, intervention compliance, and medical resource saving degree.

[0041] The adaptive communication and interaction module is used for dynamically adjusting the interaction mode according to the cognitive ability evaluation result of the old people, the interaction mode includes speed / dialect adaptation of voice interaction, font size / color contrast adjustment of interface display, multi-modal confirmation (vibration, voice, and three-mode confirmation) of instruction feedback, and provides a real-time communication channel for the old people and family members or medical staff, and is also used for synchronizing the hierarchical intervention scheme to the family members or medical staff. The cognitive ability evaluation includes memory, executive function, language ability, visual spatial ability, and calculation ability, and the interaction level is automatically matched according to the evaluation result.

[0042] The dynamic knowledge graph module further includes a pre-warning triggering sub-module, which is used for combining the association edge weight between nodes and the node attribute combination features as a pre-warning triggering condition, specifically including: presetting a multi-level pre-warning threshold, setting a basic weight threshold for different types of association edges, and setting a dynamic threshold for node attribute combination features; real-time monitoring of the association edge weight change, triggering a pre-warning when the association edge weight exceeds the basic threshold, pushing the intervention scheme to the medical staff front-end module through the adaptive communication and interaction module; identifying the node attribute combination abnormal mode, automatically activating the three-level emergency coordination when the combination features are detected to exceed the threshold, and obtaining the real-time location, sending a prompt to the old people through the adaptive communication and interaction module, and sending the risk level, associated node feature screenshot, and positioning information to the preset emergency contact person.

[0043] The above-mentioned is only an embodiment of the present application, and the specific technical solutions and / or common knowledge of the scheme are not described in detail. It should be noted that for those skilled in the art, without departing from the technical solutions of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application, which will not affect the effect and practicality of the present application. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. An artificial intelligence-based personalized health assistant management system, characterized by, The application relates to a health management system for the elderly, comprising: a multi-modal perception module for collecting physiological parameters, environmental information and behavior capture information of the elderly; an intelligent preprocessing module for desensitizing and cleaning original data collected by the multi-modal perception module through a federal learning framework, filtering motion artifact interference by adopting a time sequence network with an attention mechanism, and extracting physiological parameter features, environmental information features and behavior features; a dynamic knowledge graph module for constructing an exclusive health knowledge graph for the elderly based on the extracted features, containing basic nodes, dynamic nodes and associated edges, and updating node attributes and associated strengths through incremental learning; a personalized reasoning engine module for constructing a health management strategy generation model based on deep reinforcement learning, with the input being a real-time state vector of the knowledge graph and the output being a hierarchical intervention scheme; the hierarchical intervention scheme comprises a first-level prevention suggestion, a second-level early warning intervention and a third-level emergency coordination; an adaptive communication and interaction module for dynamically adjusting the interaction mode according to the cognitive ability evaluation result of the elderly, providing a real-time communication channel for the elderly and family members or medical staff, and synchronizing the hierarchical intervention scheme to the family members or medical staff. 2.The artificial intelligence-based personalized health assistant management system according to claim 1, characterized in that: the basic nodes comprise static information, and the static information comprises age, gender and past medical history; the dynamic nodes comprise the extracted features, and the extracted features comprise the extracted physiological parameter features, environmental information features and behavior features; the associated edges comprise the associated strengths between the features. 3.The AI-based personalized health assistant management system according to claim 1, characterized in that, the dynamic knowledge graph module comprises: a feature mapping submodule for converting the extracted features into entities and attributes of the graph; an updating submodule for determining an updating period and updating the dynamic nodes and associated edges of the graph based on newly extracted features. 4.The artificial intelligence-based personalized health assistant management system according to claim 1, wherein, the incremental learning comprises: an entity alignment submodule for automatically associating the same type of health events by calculating the semantic similarity between newly collected data and existing nodes through a graph attention network; a relationship reasoning submodule for updating the conditional probability of the associated edges based on a Bayesian network; a decay mechanism submodule for performing weight decay on historical data exceeding a preset time window. 5.The AI-based personalized health assistant management system according to claim 4, characterized in that: The incremental learning further comprises an associated node attribute updating submodule for automatically updating the attributes of the associated nodes based on the real-time updated node attributes. 6.The artificial intelligence-based personalized health assistant management system according to claim 1, wherein: The deep reinforcement learning adopts a double-delay deep deterministic policy gradient algorithm; the state space contains a plurality of feature vectors, the action space contains a combination strategy of basic intervention schemes, and the reward function is a composite index.

7. The artificial intelligence-based personalized health assistant management system according to claim 1, wherein: The cognitive ability evaluation comprises memory, executive function, language ability, visual spatial ability and calculation ability, and the evaluation result is used to automatically match an interaction level. 8.The artificial intelligence-based personalized health assistant management system according to claim 1, wherein, The dynamic knowledge graph module further comprises a warning trigger submodule for combining the associated edge weights and node attribute combination features between nodes as a warning trigger condition, specifically comprising: presetting a plurality of early warning thresholds, setting a basic weight threshold for different types of associated edges, and setting a dynamic threshold for the node attribute combination features; real-time monitoring of the associated edge weight changes, triggering a warning when the associated edge weight exceeds the basic threshold, and pushing the intervention scheme to a medical staff front-end module through the adaptive communication and interaction module. Identify node attribute combination anomaly mode, when detecting that the combination characteristics exceed the threshold value, automatically activate the three-level emergency coordination, and obtain the real-time position, send a prompt to the old people through the adaptive communication and interaction module, and send the preset emergency contact person containing the risk level, the associated node feature screenshot and the positioning information.

Images