System and method for evaluating independent living ability of old people

Abstract

The invention provides a system and a method for evaluating the independent living ability of old people. The method comprises the following steps: acquiring multi-modal original data (including physiology, behavior and interaction data) of target old people; performing modal coding, query vector interactive fusion (including missing information completion) and feature compression on the multi-modal original data to obtain a standardized feature vector; a first-layer model evaluates behavior and cognitive ability indexes, and a second-layer model combines individual background information to predict a living ability grade and a degradation risk; optimizing the first-layer model through a feature feedback mechanism; and finally, outputting and visualizing a structured evaluation result. The method breaks through the limitation of traditional manual evaluation, achieves the deep fusion of multi-modal data, improves the objectivity, robustness and interpretability of evaluation, and can meet the precise evaluation demands of home care, community monitoring and smart medical scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of health monitoring and intelligent assessment of the elderly, and in particular to an independent living ability assessment system and method for the elderly. BACKGROUND

[0002] With the intensification of global population aging, the scientific assessment of the independent living ability (including activities of daily living ADL and instrumental activities of daily living IADL) of the elderly has become the core basis for the allocation of pension service resources, the development of care programs, and the early warning of disability risks. Traditional assessment methods rely on medical personnel or caregivers to conduct manual judgments based on standardized questionnaires (such as the Barthel Index and the Lawton IADL Scale). The behavior performance of the elderly, such as "dressing, eating, and moving", needs to be recorded through on-site observation or family feedback. The following significant defects exist:

[0003] Strong subjectivity: the assessment results depend on the experience of the assessors, and the scoring difference of the same behavior by different assessors can reach 10%-20%;

[0004] Low frequency and poor real-time performance: usually assessed once a month or once a quarter, unable to capture the dynamic changes in ability (such as rapid degradation of ability caused by sudden illness);

[0005] Single data: only relying on manually recorded behavior results, lacking multi-dimensional data support such as physiological state (such as heart rate fluctuation) and environmental interaction (such as device operation).

[0006] In recent years, with the development of sensing technology, some studies have attempted to achieve automatic assessment through multi-source data, for example:

[0007] Document 1 "Detecting Anomalies in Daily Activity Routines of Older Persons in Single Resident Smart Homes" (Journal of Medical Internet Research-Aging, 2022) proposes to use environmental sensors such as door magnets and pressure pads to monitor the activity frequency of the elderly, and to judge the living ability through a rule engine. However, it can only identify "whether there is activity", and cannot quantify "activity quality" (such as the stability of getting up movements), and the multi-modal data is not modeled collaboratively;

[0008] Document 2 "MICAR: multi-inhabitant context-aware activity recognition in home environments" (Distributed and Parallel Databases, 2024) collects action data through wearable devices, identifies daily behaviors through machine learning, but the evaluation accuracy decreases by more than 30% when data is missing (such as gait data loss caused by device power failure), and the model output is only "behavior category", which cannot be directly mapped to the level of life ability.

[0009] The prior art generally has the problems of weak multi-modal fusion capability (each modal data is analyzed independently, and complementary information is not mined), poor robustness of missing data (single modal missing affects the overall evaluation), poor result interpretability (cannot locate the specific reason for ability degradation, such as "ability decline is due to slow reaction or uncoordinated action"), and unstructured output (difficult to interface with care intervention processes), which restricts its popularization and application in actual pension scenarios. SUMMARY

[0010] The purpose of the present application is to overcome the above-mentioned defects of the prior art, and to provide an independent living ability evaluation system and method for the elderly, which realizes the objectivity, automation, robustness and interpretability of evaluation.

[0011] In order to achieve the above-mentioned purpose, the present application is realized by the following technical solutions:

[0012] In a first aspect, the present application provides an independent living ability evaluation method for the elderly, comprising the following steps:

[0013] Collecting multi-modal raw data of the target elderly person, the multi-modal raw data comprising at least two of physiological data, behavior data and interaction data;

[0014] Processing the multi-modal raw data, including modal encoding, query vector interaction fusion and missing information completion, and feature compression, to obtain a standardized feature vector;

[0015] Based on the standardized feature vector, hierarchical analysis is performed, a first layer model evaluates the behavior and cognitive ability indicators of the target elderly person, and a second layer model combines individual background information and the behavior and cognitive ability indicators to predict the life ability level and ability degradation risk;

[0016] The prediction result of the second layer model is fed back to the first layer model through a feature feedback mechanism, and the feature weight and evaluation strategy of the first layer model are adjusted;

[0017] Output the life ability level, the ability degradation risk, and the structured assessment results of the behavior and cognitive ability indicators, and perform visual display.

[0018] Further as an improvement of the technical scheme of the present application, the multi-modal raw data of the target elderly person specifically includes:

[0019] The multi-modal raw data is collected by at least two sensing devices, including at least two of a wearable device (smart bracelet, smart watch), an environmental sensor (door magnet, pressure pad, temperature and humidity sensor), a video device (infrared camera, panoramic camera), a virtual reality device, and a smart home terminal (smart mattress, voice assistant);

[0020] The physiological data includes at least one of heart rate, blood oxygen saturation, blood pressure, and respiratory rate, the behavior data includes at least one of gait features (step length, step frequency), action postures (getting up, dressing, eating actions), and daily activity trajectories (kitchen, bedroom stay time), and the interaction data includes at least one of voice interaction content (demand instructions, dialogue logic) and device operation records (turning on and off lights, temperature adjustment operations);

[0021] Real-time collection (sampling frequency 1-10 Hz) and periodic collection (period 5 minutes-2 hours) are supported, and the timestamps of the multi-modal raw data are associated with the unique identifier of the target elderly person (such as the last 6 digits of the ID card + device number).

[0022] Further as an improvement of the technical scheme of the present application, the processing of the multi-modal raw data specifically includes:

[0023] The multi-modal raw data is synchronously received by a modal receiving unit, and a timestamp alignment algorithm (such as dynamic time warping (DTW)) is used to realize multi-channel data timing consistency calibration;

[0024] Each modal raw data is encoded by a modal encoding unit: visual modal data extracts dynamic behavior features through a convolutional neural network (ResNet50) + a time series network (LSTM), action modal data extracts behavior sequence features through a Transformer encoder, and physiological modal data extracts physiological state features through a fast Fourier transform (FFT) + a statistical quantity calculation (mean, variance, peak), to obtain a representation vector (dimension 256-1024) of each modal;

[0025] A learnable query vector (dimension 256) is constructed by a query interaction unit, and an attention mechanism (such as multi-head self-attention) is used to interact with each of the representation vectors, to realize cross-modal feature alignment and key ability related feature extraction;

[0026] The feature compression unit reduces the dimension of the interactive fused features to obtain the standardized feature vector with a dimension of 64-128;

[0027] The mapping output unit formats the standardized feature vector, binds the collection timestamp, target elderly identifier, and modal validity label (0 represents data missing, and 1 represents data valid), and generates input data for hierarchical analysis.

[0028] Further as an improvement of the technical scheme of the present application, before the query interaction unit constructs the learnable query vector for interactive operation, it further includes:

[0029] The query vector is pre-trained in multiple modalities, and the pre-training task includes a combination of a modal alignment task, a modal matching discrimination task, and a behavior semantic modeling task;

[0030] The modal alignment task adopts a contrast learning strategy, maps the representation vectors of different modalities to the same feature space, and makes the cosine similarity of the multi-modal vectors corresponding to the same behavior ≥0.8;

[0031] The modal matching discrimination task constructs a binary classification model (loss function is cross-entropy), trains to identify whether the input multi-modal data belongs to the same target elderly, and the accuracy is ≥90%;

[0032] The behavior semantic modeling task adopts a Seq2Seq model, generates a structured description of the behavior ability of the target elderly based on the fused features (such as “independent standing up 3 times a day, average standing up time 2.5 seconds, no action jam”).

[0033] Further as an improvement of the technical scheme of the present application, the feature compression unit reduces the dimension of the interactive fused features, specifically adopts at least one of AutoEncoder (hidden layer dimension 128-64-128), multi-layer perceptron (MLP, 3-layer structure: 1024→512→64), principal component analysis (PCA, retaining principal components with cumulative variance contribution rate ≥95%), t-distributed stochastic neighbor embedding (t-SNE), or sparse representation method (L1 regularization), retains key features related to life ability assessment (such as action completion time, heart rate fluctuation amplitude), and removes redundant dimensions (such as environmental temperature and humidity fluctuation noise).

[0034] Further as an improvement of the technical scheme of the present application, the first layer model evaluates the behavior and cognitive ability indicators of the target elderly, specifically including:

[0035] The first layer model integrates a time sequence modeling module (such as a GRU network) to extract time correlation features of the standardized feature vectors within a continuous collection period (1 hour-24 hours);

[0036] At least one of a logistic regression, a light weight regression network (2-layer MLP: 64→32→output) or a decision tree model (XGBoost) is adopted to output the behavior and cognitive ability indicators;

[0037] The behavior and cognitive ability indicators include a reaction speed score (0-10 points, the higher the score, the faster the reaction), an execution stability score (0-10 points, the higher the score, the more stable the action), a concentration score (0-10 points, the higher the score, the more concentrated the attention), and a motion coordination score (0-10 points, the higher the score, the better the coordination), and the calculation error of each indicator is ≤0.5 points.

[0038] Further as an improvement of the technical scheme of the present application, the second layer model combines individual background information and the behavior and cognitive ability indicators to predict a life ability grade and an ability degradation risk, specifically including:

[0039] The individual background information includes age (60-120 years old), a history of underlying diseases (binary labels such as hypertension and diabetes), historical ability evaluation records (scores in the past 3 months), and living environment characteristics (classification labels such as living alone or in a group and having or not having a handrail), which are converted into vector features through an embedding layer (dimension 32-64);

[0040] The behavior and cognitive ability indicators (4-6 dimensions, each dimension 0-10 points) and the individual background information vector are spliced to obtain input features (dimension 64-128);

[0041] The second layer model adopts at least one of a multi-layer perception (MLP, 4-layer structure: 128→64→32→output), a graph neural network (GNN, based on a disease-ability correlation graph) or a gradient boosting tree model (LightGBM) to learn a nonlinear correlation between the behavior and cognitive ability indicators and the life ability;

[0042] The output life ability grade is divided into 0-5 levels (0 level: complete disability, 5 level: complete independence), the ability degradation risk includes a risk index (0-1, the larger the value, the higher the risk) and a risk trend label (stable: score fluctuation ≤0.5 points in the past 1 month; mild decline: score decrease 0.5-1 points; rapid degradation: score decrease >1 point), and the prediction accuracy is ≥85%.

[0043] Further as an improvement of the technical scheme of the application, the prediction result of the second layer model is fed back to the first layer model through a feature feedback mechanism, specifically including:

[0044] Based on the life ability level and ability degradation risk output by the second layer model, SHAP value analysis is used to determine the behavior and cognitive ability dimensions that have a significant impact on the evaluation result (for example, in the "rapid degradation" risk, the absolute value of the SHAP of the reaction speed dimension is greater than or equal to 0.3);

[0045] According to the significant dimension, the weight coefficient of the corresponding feature in the first layer model is adjusted (for example, the weight of the reaction speed related feature is increased by 20-50%), and the sensitivity of the first layer model to the significant dimension change is enhanced (for example, a decrease of 0.5 in reaction speed can trigger a risk warning);

[0046] Every 7-30 days, repeat the feedback adjustment based on new evaluation data, and use the incremental learning strategy to update the parameters of the first layer model, so that the evaluation accuracy of the model remains greater than or equal to 85% in long-term operation, and the generalization ability is improved by 10-15%.

[0047] The second aspect of the application provides an independent living ability evaluation system for the elderly, comprising:

[0048] A data acquisition unit is configured to acquire multi-modal raw data of a target elderly person, wherein the multi-modal raw data includes at least two of physiological data, behavior data and interaction data.

[0049] A data processing unit includes a modal receiving module, a modal encoding module, a query interaction module, a feature compression module and a mapping output module; the modal receiving module is configured to align the multi-channel data time sequence, the modal encoding module is configured to independently encode each modal data, the query interaction module is configured to cross-modal fusion and missing data completion, the feature compression module is configured to reduce the dimension of the feature, and the mapping output module is configured to generate standardized input data.

[0050] A data analysis unit includes a first layer model module, a second layer model module and a feature feedback module; the first layer model module is configured to evaluate the behavior and cognitive ability indicators of the target elderly person, the second layer model module is configured to predict the life ability level and ability degradation risk in combination with individual background information and the behavior and cognitive ability indicators, and the feature feedback module is configured to feed back the prediction result of the second layer model module to the first layer model module to adjust the feature weight and the evaluation strategy.

[0051] A visual output unit is configured to output and visualize the structured assessment results, including the life ability level, the ability degradation risk and the behavior and cognitive ability indicators, in the form of an ability radar chart (dimension score distribution), a risk trend line chart (risk change in the last 1-3 months) and numerical labels (ability level and risk index), and support mobile APP and desktop platform access.

[0052] The third aspect of the application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to realize the method for evaluating the independent living ability of the elderly person.

[0053] Compared with the prior art, the application has the following beneficial effects:

[0054] Objective evaluation: fully based on automatic evaluation of multi-modal data, avoiding artificial subjective bias, and the consistency of the evaluation results (repeated evaluation at different time points) is greater than or equal to 90%;

[0055] Strong multi-modal fusion capability: complementary information of each mode is mined through a query interaction mechanism, such as "heart rate data assisting in verifying motion stability", and the evaluation accuracy is improved by 25%-35% compared with a single mode;

[0056] High robustness of missing data: pre-training tasks realize missing information completion, and the evaluation accuracy decreases by less than or equal to 5% when a single mode is missing, which is much better than the prior art (decreases by 30%+);

[0057] Good result interpretability: hierarchical modeling locates specific dimensions of ability degradation (such as "ability decreases due to slow reaction speed"), and supports precise positioning of care intervention;

[0058] Output structure: the generated life ability level and risk index can be directly connected to the old-age service process (such as triggering an emergency care response when the risk index is greater than 0.7);

[0059] Long-term stability: the feature feedback mechanism makes the evaluation accuracy of the model remain greater than or equal to 85% in a 6-month operation cycle without significant decay.

[0060] The application can be widely applied to home-based care (remote monitoring by family members), community-based care (resource optimization), geriatric hospitals (rehabilitation effect evaluation) and other scenarios, and provides an efficient and accurate technical solution for the evaluation of the life ability of the elderly, which has important social value and market promotion prospects. BRIEF DESCRIPTION OF DRAWINGS

[0061] Other features, objects and advantages of the application will become more apparent after reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0062] Figure 1 A schematic diagram of the overall structure of an independent living ability assessment system for the elderly according to an embodiment of the present application

[0063] Figure 2 A schematic diagram of the internal structure of a data processing unit according to an embodiment of the present application

[0064] Figure 3 A schematic diagram of the internal structure of a data analysis unit according to an embodiment of the present application

[0065] Figure 4 A schematic diagram of the flow of an independent living ability assessment method for the elderly according to an embodiment of the present application

[0066] Figure 5 A schematic diagram of the composition of an electronic device according to an embodiment of the present application DETAILED DESCRIPTION

[0067] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments, which are intended to explain the present application and are not intended to limit the present application.

[0068] Note that all directional references (e.g., upper, lower, left, right, front, rear, top, bottom, under, above, etc.) are in relation to the exemplary embodiment of the present application as shown in the drawings and are used only to illustrate the present application and are not intended to limit the present application unless clearly indicated by the claims.

[0069] In the present application, unless otherwise clearly specified and limited, the term “connection” should be understood broadly, for example, “connection” can be fixed connection, can be detachable connection, or can be integrated; can be mechanical connection, can be electrical connection; can be direct connection, can be indirect connection through an intermediate medium, can be internal communication of two elements or interaction relationship between two elements, unless otherwise clearly specified and limited. For those skilled in the art, the specific meaning of the above-mentioned term in the present application can be understood according to the specific circumstances.

[0070] In addition, in the present application, the description such as “first”, “second” and the like is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features limited by “first”, “second” can explicitly or implicitly include at least one of the features; in addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the fact that a person skilled in the art can realize it, when the combination of technical solutions appears contradictory or cannot be realized, it should be considered that the combination of technical solutions does not exist, nor within the scope of protection required by the present application.

[0071] The application will be described in further detail below with reference to the drawings.

[0072] Referring to Figure 4 In a first aspect, the application provides an independent living ability assessment method for the elderly, comprising the following steps:

[0073] Collecting multi-modal raw data of the target elderly person, the multi-modal raw data comprising at least two of physiological data, behavior data and interaction data;

[0074] Processing the multi-modal raw data, including modal encoding, query vector interaction fusion and missing information completion, feature compression, to obtain a standardized feature vector;

[0075] Performing hierarchical analysis based on the standardized feature vector, a first layer model evaluating the behavior and cognitive ability indicators of the target elderly person, and a second layer model combining individual background information and the behavior and cognitive ability indicators to predict the living ability grade and the ability degradation risk;

[0076] Feeding back the prediction result of the second layer model to the first layer model through a feature feedback mechanism to adjust the feature weight and evaluation strategy of the first layer model;

[0077] Outputting the structured evaluation results of the living ability grade, the ability degradation risk and the behavior and cognitive ability indicators, and performing visual display.

[0078] The core of the application realizes the automatic and objective evaluation of the independent living ability of the elderly through the whole process of "multi-modal data collection, intelligent fusion processing, hierarchical modeling analysis, feedback optimization and visual output", and the specific process is as follows:

[0079] Multi-modal raw data collection: through video equipment (infrared camera, panoramic camera), virtual reality equipment (for cognitive ability test), environmental sensors (door magnet, pressure mat, temperature and humidity sensor) and wearable devices (smart bracelet, wearable camera), multi-modal raw data of the target elderly person is collected, at least two of physiological data (heart rate, blood oxygen saturation, blood pressure), behavior data (gait characteristics, action posture, daily activity trajectory), interaction data (voice interaction content, smart home operation record) are covered, and a data basis fully reflecting the life state of the elderly is constructed;

[0080] Intelligent processing of multi-modal data:

[0081] Modal encoding: adapt encoding methods to different modal data characteristics, map heterogeneous data to a unified feature space (e.g., visual data extracts dynamic behavior features through "ResNet50+LSTM", action data extracts sequence features through Transformer encoder, physiological data extracts state features through "FFT+statistical quantity calculation");

[0082] Query vector interaction fusion: construct a learnable query vector (dimension 256), use multi-head self-attention mechanism to interact with each modal encoding representation vector, realize cross-modal feature alignment; for data missing scenarios (e.g., device power failure leading to gait data missing), complete missing information based on inter-modal correlation;

[0083] Feature compression: remove environmental noise and other redundant dimensions through dimension reduction algorithms such as AutoEncoder and PCA, generate 64-128 dimensional standardized feature vectors;

[0084] Hierarchical modeling analysis:

[0085] First layer model: based on standardized feature vectors, integrate GRU time sequence module to extract time correlation features of continuous collection period (1-24 hours), use logistic regression, lightweight MLP, etc. Model, output behavior and cognitive ability indicators (0-10 quantization) such as reaction speed, execution stability, attention concentration, and action coordination;

[0086] Second layer model: combine the ability indicators output by the first layer and individual background information (age, medical history, historical evaluation records, living environment characteristics), use 4-layer MLP, GNN or LightGBM model to predict life ability level (0-5 levels, 0 completely disabled, 5 completely independent) and ability degradation risk (risk index 0-1, trend label "stable / mild decline / rapid degradation");

[0087] Feature feedback optimization: based on the prediction results of the second layer model, use SHAP value analysis to locate the key ability dimensions that affect the evaluation results (e.g., reaction speed is the key dimension in "rapid degradation" risk), adjust the feature weights of the first layer model in the opposite direction (e.g., reaction speed related feature weight increases by 20%-50%), and enhance the sensitivity of the model to key dimension changes;

[0088] Visual output: structured results such as life ability level, ability degradation risk, and behavior and cognitive ability indicators are displayed through ability radar chart, risk trend line chart, numerical label card (green low risk, yellow medium risk, red high risk), etc. Form, support mobile APP, community care platform and hospital system access.

[0089] The embodiment has the following beneficial effects:

[0090] Breakthrough traditional evaluation limitations: solve the problems of "strong subjectivity (10%-20% difference in scoring by different evaluators), low frequency (once a month), and poor real-time performance" of traditional manual questionnaires, realize 24-hour continuous monitoring and automatic evaluation, and the consistency of evaluation results (repeated evaluation at different time points) is ≥90%;

[0091] Outstanding multi-modal fusion capability: excavate the complementary value of physiological, behavioral, and interactive data (such as heart rate data verifying the authenticity of motion stability), avoid one-sidedness of single modal data, and improve evaluation accuracy by 25%-35% compared with single modal;

[0092] Robustness and interpretability: missing information completion mechanism makes the evaluation accuracy decrease ≤5% when single modal is missing (much better than the existing technology with a decrease of 30%+), hierarchical modeling can locate the specific dimension of capability degradation (such as "capability degradation only due to low attention concentration"), and provide basis for precise care intervention;

[0093] Strong practicality: structured output can directly interface with elderly care service processes (such as risk index >0.7 triggering emergency care response), visual form reduces the understanding threshold of non-professionals (such as family members), and adapts to home care, community monitoring, and elderly medical care in multiple scenarios.

[0094] In some embodiments, the multi-modal raw data of the target elderly person is collected, specifically including:

[0095] The multi-modal raw data is collected by at least two sensing devices, including at least two of wearable devices (smart bracelet, smart watch), environmental sensors (door magnet, pressure mat, temperature and humidity sensor), video devices (infrared camera, panoramic camera), virtual reality devices, and smart home terminals (smart mattress, voice assistant);

[0096] The physiological data includes at least one of heart rate, blood oxygen saturation, blood pressure, and respiratory rate, the behavioral data includes at least one of gait features (step length, step frequency), motion posture (getting up, dressing, eating motion), and daily activity trajectory (kitchen, bedroom stay time), and the interactive data includes at least one of voice interaction content (demand instruction, dialogue logic) and device operation record (turning on and off lights, temperature adjustment operation);

[0097] Support real-time collection (sampling frequency 1-10Hz) and periodic collection (period 5 minutes-2 hours), and associate the time stamp of the multi-modal raw data with the unique identifier of the target elderly person (such as the last 6 digits of the ID card + device number).

[0098] The application ensures the integrity, timeliness and traceability of the data by "diversified selection of sensing devices -> classification data collection strategy -> standardized time sequence and identification association". In specific implementation, the selection and deployment of sensing devices:

[0099] Home scene: preferentially deploy non-intrusive devices, including smart bracelet (collecting heart rate, blood oxygen, action posture, sampling frequency 1-10Hz), door magnetic sensor (recording door opening and closing times, triggering collection), smart mattress (collecting sleep duration and turning frequency, sampling period 1 minute), voice assistant (recording voice instructions, wake-up collection);

[0100] Community / medical scene: supplement infrared camera (collecting gait features, frame rate 15fps), panoramic camera (collecting complex actions such as dressing and eating, frame rate 10fps), virtual reality device (for cognitive ability test, such as reaction speed training), covering public areas and professional assessment needs;

[0101] Data classification collection strategy:

[0102] Physiological data (heart rate, blood oxygen, blood pressure): real-time collection (sampling frequency 1-10Hz) is adopted to ensure dynamic capture of body state changes (such as heart rate rising when getting up);

[0103] Behavior data (gait, action posture, activity trajectory): periodic collection (period 5 minutes-2 hours) is adopted to balance data integrity and storage pressure;

[0104] Interaction data (voice instructions, device operation): wake-up / trigger collection (only when the user triggers the voice assistant or operates the device, record), reduce redundant data;

[0105] Data time sequence and identification association:

[0106] Time sequence alignment: all collected data are bound with accurate time stamp (UTC format, accuracy ±1ms), through NTP protocol to realize multi-device time synchronization, to ensure that "getting up action at 10:00:00" and "heart rate data at 10:00:00" are accurately associated;

[0107] Individual identification: bind a unique identifier for each data (rule: last 6 digits of ID card + device type code + last 4 digits of device MAC, such as "123456-W-ABCD", "W" represents smart bracelet), to avoid multi-user data confusion;

[0108] Storage management: data is transmitted to the edge computing node through 5G / Wi-Fi, stored in the InfluxDB time series database, and organized in a three-level directory of "elderly ID-equipment type-time" (such as "123456 / W / 2024-05-01") for subsequent query and call.

[0109] This embodiment has the following beneficial effects:

[0110] Data source comprehensiveness: covering the three dimensions of "physiology-behavior-interaction", avoiding one-sidedness of a single device / data type (such as heart rate data of a smart bracelet can verify whether the "getting up action" collected by a camera is abnormal);

[0111] Flexible collection strategy: real-time collection meets the needs of physiological data dynamic monitoring, periodic collection reduces the storage pressure of behavior data, wake-up collection reduces the redundancy of interaction data, and adapts to the characteristics of different data;

[0112] Data traceability: timestamp ensures the time sequence consistency of multi-modal data (lays the foundation for subsequent cross-modal fusion), unique identifier realizes accurate association of individual data, and avoids data confusion in multi-user scenarios such as communities;

[0113] Strong scene adaptability: device selection considers "non-invasive" and "complementary functions", which can be deployed in home, community, and medical scenarios without large-scale environmental modification, reducing the cost of landing.

[0114] In some embodiments, the processing of the multi-modal raw data specifically includes:

[0115] The multi-modal raw data is synchronously received by a modal receiving unit, and a timestamp alignment algorithm (such as dynamic time warping, DTW) is used to realize multi-channel data time sequence consistency calibration;

[0116] Each modal raw data is encoded by a modal encoding unit: visual modal data extracts dynamic behavior features through "convolutional neural network (ResNet50) + time sequence network (LSTM)", action modal data extracts behavior sequence features through a Transformer encoder, and physiological modal data extracts physiological state features through "fast Fourier transform (FFT) + statistical quantity calculation (mean, variance, peak)", obtaining a representation vector (dimension 256-1024) of each modal;

[0117] A learnable query vector (dimension 256) is constructed by a query interaction unit, and an attention mechanism (such as multi-head self-attention) is used to interact with each representation vector, realizing cross-modal feature alignment and key ability related feature extraction;

[0118] The standardized feature vector with a dimension of 64-128 is obtained by dimension reduction processing of the interactive fused features through a feature compression unit;

[0119] The standardized feature vector is format normalized through a mapping output unit, a collection timestamp, a target elderly identifier and a modal validity label (0 represents data missing, and 1 represents data valid) are bound, and input data for hierarchical analysis is generated.

[0120] The present application converts heterogeneous raw data into high-quality feature vectors suitable for hierarchical analysis through the five-step process of “time alignment → modal encoding → cross-modal fusion → feature compression → format standardization”, as follows:

[0121] Time consistency calibration:

[0122] After the modal receiving unit receives the multi-device data, the timestamp of the highest sampling frequency device (such as a pressure pad 10Hz) is taken as the reference, the dynamic time warping (DTW) algorithm is used to interpolate and align the data of the low sampling frequency device (such as a smart bracelet 1Hz), the time sequence deviation caused by the sampling period difference of the multi-device is eliminated, and the deviation after alignment is ≤0.1 seconds;

[0123] Modal independent encoding:

[0124] Visual modal (video data): image space features (such as human body skeleton key points) are extracted through ResNet50, dynamic gait features within 30 seconds are extracted by combining LSTM (time step 30), and a 512-dimensional representation vector is output;

[0125] Action modal (wearable device acceleration / gyroscope data): behavior sequence rules are extracted through a 6-layer Transformer encoder (attention head number 8), and a 256-dimensional representation vector is output;

[0126] Physiological modal (heart rate / blood oxygen data): the original data is converted to the 0.1-1Hz frequency domain (corresponding to the normal heart rate range) through FFT, 10 statistical quantities such as mean, variance and peak value are calculated, and a 128-dimensional representation vector is mapped through a fully connected layer;

[0127] Cross-modal fusion and missing completion:

[0128] A 256-dimensional learnable query vector is constructed (parameters are optimized through multi-modal pre-training), and a multi-head self-attention mechanism is used to calculate the attention weights of the query vector and the representation vectors of each modality (the higher the weight, the greater the contribution of the modality to the ability assessment); for the modality whose modality effectiveness label is 0 (data missing), complementary features are extracted from other valid modalities based on the attention weights to complete the completion (for example, when physiological data is missing, the physiological state features are completed through the association of "visual gait + action posture");

[0129] Feature dimension compression:

[0130] An AutoEncoder (input dimension 896 = 512+256+128, encoder structure 896→512→256→128) or PCA (retaining principal components with cumulative variance contribution rate ≥95%) is used to remove redundant dimensions such as environmental humidity fluctuations and device acquisition noise, and generate 64-128-dimensional standardized feature vectors;

[0131] Format standardization:

[0132] The mapping output unit binds the "acquisition timestamp (UTC format), target elderly person identifier, modality effectiveness label (0 = data missing, 1 = data valid)" for the standardized feature vector, generates a structured format "[identifier, timestamp, feature vector, modality label]", and ensures that the hierarchical analysis unit can clearly identify the feature source and integrity.

[0133] The embodiment has the following beneficial effects:

[0134] Time series deviation elimination: the DTW algorithm solves the data misplacement problem caused by the difference in sampling period of multiple devices (such as "10:00:00.5 getting up action" and "10:00:01 heart rate sudden rise" correct association), providing a reliable time series basis for cross-modal fusion;

[0135] Modality encoding is targeted: different encoding methods are selected according to the characteristics of different modal data to avoid feature loss caused by "one-size-fits-all" encoding (such as the Transformer encoder, which improves the encoding accuracy of action sequences by 15% compared with traditional CNN);

[0136] Efficient cross-modal fusion: the attention mechanism can adaptively allocate weights to each modality (such as when physiological modalities are missing, automatically increasing the weight of action modalities), and the missing information completion capability makes the model maintain stable performance when the data is incomplete;

[0137] High feature quality: feature compression removes redundant dimensions, reduces subsequent model calculation (inference speed improves by 30%+) while retaining key features related to life ability assessment (compressed features have a correlation coefficient of 0.8 or more with ability indicators);

[0138] Interface friendliness: standardized format allows feature vectors to directly interface with the input interface of the layered analysis unit, without additional data conversion, improving system modularity and maintainability.

[0139] In some embodiments, the query interaction unit further includes, before constructing the learnable query vector for interactive operation:

[0140] Multi-modal pre-training is performed on the query vector, and the pre-training task includes a combination of modal alignment task, modal matching discrimination task, and behavior semantic modeling task;

[0141] The modal alignment task uses a contrast learning strategy to map the representation vectors of different modalities to the same feature space, so that the cosine similarity of the multi-modal vectors corresponding to the same behavior is greater than or equal to 0.8;

[0142] The modal matching discrimination task constructs a binary classification model (loss function is cross-entropy) to identify whether the input multi-modal data belongs to the same target elderly person, with an accuracy of greater than or equal to 90%;

[0143] The behavior semantic modeling task uses a Seq2Seq model to generate a structured description of the behavior ability of the target elderly person based on the fused features (e.g., "independent standing up 3 times a day, average standing up time 2.5 seconds, no motion lag").

[0144] The present application improves the cross-modal processing capability of the query vector through the cooperative training of the three tasks of "modal alignment + modal matching discrimination + behavior semantic modeling", as follows:

[0145] Modal alignment task (spatial alignment):

[0146] A "positive sample - negative sample" pair is constructed using a contrast learning strategy - the positive sample is different modal data of the same behavior of the same elderly person (e.g., "visual data + motion data of the standing up action of Elder A"), and the negative sample is modal data of the same behavior of different elderly people (e.g., "visual data of the standing up action of Elder A + motion data of the standing up action of Elder B"); the query vector is trained to map the multi-modal representation vectors of the positive sample to the same feature space (cosine similarity is greater than or equal to 0.8), and to pull apart the vectors of the negative sample (cosine similarity is less than or equal to 0.3), achieving cross-modal spatial alignment;

[0147] Modal matching discrimination task (relevance judgment):

[0148] A binary classification model is constructed (input: query vector and multi-modal fusion feature, output: "match 1 / mismatch 0"), and the training data includes matching samples (multi-modal data of the same elderly person, such as "old man A's heart rate data + old man A's gait data") and mismatching samples (multi-modal data of different elderly people, such as "old man A's heart rate data + old man B's gait data"); a cross-entropy loss function is used to optimize the model parameters, so that the accuracy of the model in identifying the correlation of multi-modal data is ≥ 90%;

[0149] Behavioral semantic modeling task (semantic understanding):

[0150] A Seq2Seq model (encoder: fusion module of query interaction unit, decoder: 2-layer LSTM) is used, input multi-modal fusion feature, output structured description of behavior ability in natural language form (such as "old man B gets up independently 3 times a day, average getting up time 2.5 seconds, no action lag"); during training, the Adam optimizer and cross-entropy loss function are used to make the generated description consistent with the manual annotation ≥ 85%;

[0151] Pre-training parameter migration: the pre-trained query vector parameters are used as the initial parameters of the query interaction unit, and a small amount of fine-tuning (based on 500 scenario samples) is performed to adapt to specific application scenarios (such as adding "smart mattress data" for home scenarios).

[0152] The embodiment has the following beneficial effects:

[0153] Cross-modal alignment accuracy is improved: the modal alignment task significantly improves the consistency of different modal data in the feature space (cosine similarity ≥ 0.8), avoiding fusion bias caused by modal space differences (such as the contradiction between visual modal judgment "normal gait" and action modal judgment "abnormal gait");

[0154] Data correlation judgment is accurate: the modal matching discrimination task enables the model to accurately identify the attribution consistency of multi-modal data (accuracy ≥ 90%), avoiding the misassociation of "old man A's heart rate data" and "old man B's action data" in a multi-user scenario such as a community;

[0155] Behavioral semantic understanding ability is enhanced: the behavioral semantic modeling task enables the fusion feature to have interpretable semantic information, not only providing a "feature - semantic" mapping basis for hierarchical analysis (such as "high dimension in fusion feature → stable getting up action"), but also being directly used for visualized semantic description;

[0156] High reliability of missing data completion: the pre-trained query vector can extract complementary features from the effective modalities based on the modal alignment and matching relationship (e.g., when physiological data is missing, the physiological state features are completed based on the "visual gait - action data" alignment relationship), and the completion error is ≤10%.

[0157] In some embodiments, the feature compression unit performs dimension reduction processing on the fused features, specifically at least one of AutoEncoder (hidden layer dimension 128-64-128), multilayer perceptron (MLP, 3-layer structure: 1024→512→64), principal component analysis (PCA, retaining principal components with cumulative variance contribution rate ≥95%), t-distributed stochastic neighbor embedding (t-SNE), or sparse representation method (L1 regularization), retaining key features related to life ability assessment (such as action completion time, heart rate fluctuation amplitude), and removing redundant dimensions (such as environmental temperature and humidity fluctuation noise).

[0158] The present application balances between "retaining key ability features" and "reducing computational dimension", removes redundant information, and specifically as follows:

[0159] AutoEncoder (AutoEncoder) dimension reduction:

[0160] The "encoder - decoder" symmetric structure is adopted - the encoder input is 896-dimensional fused features (visual 512-dimensional + action 256-dimensional + physiological 128-dimensional), which is mapped to 128-dimensional latent vector (standardized feature vector) through hidden layer 896→512→256→128 (ReLU activation); the decoder reconstructs the original fused features through hidden layer 128→256→512→896 (Sigmoid activation); during training, the MSE reconstruction error is minimized to make the encoder learn to retain key features such as action completion time and heart rate fluctuation amplitude, and discard redundant dimensions such as environmental temperature and humidity noise;

[0161] MLP (multilayer perceptron) dimension reduction:

[0162] A 3-layer fully connected network (input 1024-dimensional→512-dimensional→64-dimensional, ReLU activation) is adopted, and the training target is the accuracy of the compressed feature vector in predicting the life ability level; the cross-entropy loss function is used to optimize the network parameters, so that the output 64-dimensional feature only retains key information strongly related to ability assessment (such as gait step variance, reaction speed correlation features);

[0163] PCA (principal component analysis) dimension reduction:

[0164] Standardize the fusion feature matrix (sample number N x feature dimension D), calculate the covariance matrix and solve the eigenvalue and eigenvector; select the top K eigenvectors (principal components) with cumulative variance contribution rate ≥95% after sorting the eigenvalues from large to small, and project the original D-dimensional features to K-dimensional space (K=64-128); PCA has no parameter learning, high calculation efficiency, and is suitable for real-time evaluation scenarios;

[0165] t-SNE (t-distributed Stochastic Neighbor Embedding) dimensionality reduction:

[0166] Calculate the similarity between samples in high-dimensional space through a Gaussian kernel function, and calculate the similarity between samples in low-dimensional space through a t-distribution kernel function. Minimize the KL divergence of the two, and map the high-dimensional features to a 64-128 dimensional low-dimensional space. t-SNE is good at preserving the local structure in high-dimensional space (such as samples with similar life ability levels clustering in low-dimensional space), and is suitable for feature input of the first layer of the hierarchical model (behavior and cognitive ability evaluation);

[0167] Sparse representation (L1 regularization) dimensionality reduction:

[0168] Add an L1 regularization term (loss function = reconstruction error / prediction error + λ・||W||1, λ=0.01) in the feature compression process to make the network weight W sparse (a large number of weights tend to 0); dimensions with absolute weight <1e-5 are determined as redundant dimensions and removed, and finally the key feature dimensions corresponding to the non-zero weights (64-128 dimensions) are retained.

[0169] The embodiment has the following beneficial effects:

[0170] Improved calculation efficiency: compressing high-dimensional fusion features (896-1024 dimensions) to 64-128 dimensions reduces the computational load of subsequent hierarchical models (inference time is shortened from 500ms to within 100ms), supporting the deployment of edge computing nodes (such as Raspberry Pi 4B);

[0171] Key feature preservation: through target constraints such as reconstruction error, prediction accuracy, and variance contribution rate, ensure that the correlation coefficient of the compressed features with life ability evaluation is ≥0.8, and the evaluation accuracy decreases ≤2%;

[0172] Redundant noise removal: effectively filters redundant information such as environmental temperature and humidity fluctuations, device acquisition noise (such as camera light interference, wearable device motion artifacts), etc., improving the feature signal-to-noise ratio (the signal-to-noise ratio after compression is improved by 20%);

[0173] Method flexibility: according to the scene requirements, the appropriate dimension reduction method (such as PCA suitable for scenes with high real-time requirements, and AutoEncoder suitable for scenes with high feature expression capability requirements) can be selected to enhance the scene adaptability of the system.

[0174] In some embodiments, the first layer model evaluates the behavior and cognitive ability indicators of the target elderly person, specifically including:

[0175] The first layer model integrates a time series modeling module (such as a GRU network) to extract the time correlation features of the standardized feature vectors within a continuous collection period (1 hour-24 hours);

[0176] At least one of the following is used: logistic regression, a lightweight regression network (2-layer MLP: 64→32→output), or a decision tree model (XGBoost) to output the behavior and cognitive ability indicators;

[0177] The behavior and cognitive ability indicators include reaction speed scores (0-10 points, with higher scores indicating faster reaction), execution stability scores (0-10 points, with higher scores indicating more stable motion), attention concentration scores (0-10 points, with higher scores indicating more concentrated attention), and motion coordination scores (0-10 points, with higher scores indicating better coordination), with a calculation error of ≤0.5 points for each indicator.

[0178] The present application generates interpretable intermediate evaluation results through "time series feature extraction → lightweight model modeling → multi-dimensional indicator output", as follows:

[0179] Time series correlation feature extraction:

[0180] The first layer model integrates a GRU (Gated Recurrent Unit) time series module, with the input being a sequence of standardized feature vectors in a continuous collection period (1-24 hours, sampled at 5 minutes per time, with a total of 12-288 time points); through the update gate (controls the proportion of historical information retained) and the reset gate (controls the degree of historical information ignored) of the GRU, the time correlation of the features (such as "early 8 o'clock - late 8 o'clock reaction speed trend") is captured, and a 256-dimensional time series fusion feature is output;

[0181] Lightweight model modeling:

[0182] For each behavior and cognitive ability indicator, an independent model with small computational complexity and strong interpretability is used:

[0183] Reaction speed score (0-10 points): a logistic regression model is used, based on the "voice command response time" correlation feature (response time <1 second → 10 points, 1-2 seconds → 8 points, 2-3 seconds → 5 points, >3 seconds → 0 points), to output the score;

[0184] Stability score (0-10): 2-layer MLP (input 256-dim→32-dim→1-dim) based on "standard deviation of standing motion pressure distribution" correlation feature (standard deviation <50kPa→10 points, 50-100kPa→8 points, 100-200kPa→5 points, >200kPa→0 points), output score;

[0185] Concentration score (0-10): XGBoost decision tree model based on "number of device operation errors" correlation feature (error times 0→10 points, 1-2 times→8 points, 3-5 times→5 points, >5 times→0 points), output score;

[0186] Motion coordination score (0-10): 2-layer MLP (input 256-dim→32-dim→1-dim) based on "gait step length variance" correlation feature (variance <5cm→10 points, 5-10cm→8 points, 10-20cm→5 points, >20cm→0 points), output score;

[0187] Error control of indicators:

[0188] Ensure model stability through 5-fold cross-validation (MAE difference ≤0.1 points); for the score output by the model, if the difference with the previous 3 scores is >2 points (considered abnormal, such as device failure), replace it with the average of the previous 3 scores, ensuring that the calculation error of each indicator is ≤0.5 points.

[0189] This embodiment has the following beneficial effects:

[0190] Strong interpretability: output specific behaviors and cognitive ability dimensions (such as reaction speed, motion stability), rather than abstract feature vectors, which facilitates caregivers to accurately identify the ability shortcomings of the elderly (such as "ability decline only due to low concentration, other dimensions normal");

[0191] Time series correlation capture: GRU module can extract the ability change trend in continuous periods (such as "reaction speed decreased by 1.5 points in the afternoon compared to the morning"), avoiding the one-sidedness of single-time-point data and improving the comprehensiveness of evaluation;

[0192] High computational efficiency: lightweight model (such as 2-layer MLP) with inference time ≤50ms, suitable for edge computing nodes, meeting real-time evaluation needs (evaluation frequency 1 time / 5 minutes);

[0193] Controllable error: the error of each index calculation is less than or equal to 0.5 points, which ensures the reliability of the evaluation results (for example, the difference between 8.0 points and 8.5 points can be clearly distinguished), and provides high-precision input for the second layer model.

[0194] In some embodiments, the second layer model combines individual background information and the behavior and cognitive ability index to predict the life ability level and the ability degradation risk, specifically including:

[0195] The individual background information includes age (60-120 years old), basic disease history (binary label of hypertension, diabetes, etc.), historical ability evaluation record (score in the past 3 months), living environment characteristics (classification label of living alone / group, presence / absence of handrails, etc.), which are converted into vector features through an embedding layer (dimension 32-64);

[0196] The behavior and cognitive ability index (4-6 dimensions, each dimension 0-10 points) and the individual background information vector are spliced to obtain input features (dimension 64-128);

[0197] The second layer model uses at least one of a multi-layer perception (MLP, 4-layer structure: 128→64→32→output), a graph neural network (GNN, based on a disease-ability association graph) or a gradient boosting tree model (LightGBM) to learn the nonlinear correlation between the behavior and cognitive ability index and the life ability;

[0198] The output life ability level is divided into 0-5 levels (0 level: complete disability, 5 level: complete independence), the ability degradation risk includes a risk index (0-1, the larger the value, the higher the risk) and a risk trend label (stable: score fluctuation≤0.5 points in the past 1 month; mild decline: score decrease 0.5-1 points; rapid degradation: score decrease>1 points), and the prediction accuracy is≥85%.

[0199] The present application realizes the progression from "local index" to "overall evaluation" through "background information embedding→feature splicing→nonlinear modeling→multi-dimensional output", specifically as follows:

[0200] Individual background information embedding:

[0201] Convert heterogeneous individual background information into a unified vector:

[0202] Age (60-120 years old): mapped to a 32-dimensional vector through a linear layer to ensure that the age and vector element values are monotonous (the larger the age, the larger the vector value);

[0203] History of underlying diseases (5 categories: hypertension, diabetes, etc.): One-hot encoding (e.g., "hypertension + diabetes" → [1, 1, 0, 0, 0]) and then mapped to a 32-dimensional vector through an embedding layer;

[0204] History of ability assessment records (scores in the past 3 months): Extract time-series features through a 1D-CNN (convolution kernel size 3, output channel 32), and generate a 32-dimensional vector through global average pooling;

[0205] Living environment features (residential type, auxiliary facilities, etc.): mapped to a 32-dimensional vector through an embedding matrix (9x32-dimensional);

[0206] Finally, concatenate to generate a 32-64-dimensional individual background information vector;

[0207] Input feature construction:

[0208] Concatenate the 4 behavior and cognitive ability indicators (4-dimensional, 0-10 points) output by the first layer model with the individual background information vector (32-64-dimensional) to generate a 64-128-dimensional input feature (e.g., 4-dimensional + 64-dimensional = 68-dimensional);

[0209] Non-linear association modeling:

[0210] Use a 4-layer MLP (128→64→32→output), GNN (based on "underlying disease - ability indicator" association graph), or LightGBM model to learn the complex non-linear relationship between "behavior and cognitive indicators + individual background information" and life ability:

[0211] MLP: end-to-end modeling, suitable for capturing high-order feature interactions (e.g., "age > 80 years old + hypertension + reaction speed < 5 points" superposition effect);

[0212] GNN: through graph structure to mine the association path of "underlying disease→ability indicator→life ability" (e.g., "hypertension→decreased executive stability→reduced life ability"), enhancing interpretability;

[0213] LightGBM: suitable for handling high-dimensional sparse features (e.g., time-series sparsity of historical assessment records), fast training speed;

[0214] Multi-dimensional output generation:

[0215] Life ability level: output probability distribution of 0-5 levels through softmax activation function, take the highest probability level as the final result (0 level completely disabled, 1 level severely disabled, 2 level moderately disabled, 3 level mildly disabled, 4 level basically independent, 5 level completely independent);

[0216] Capability degradation risk: output a risk index of 0-1 through a sigmoid activation function (calculation basis: base risk 30% + capability risk 50% + trend risk 20%); combined with the trend of the change in the capability index in the past 1 month (mean difference), generate a trend label of “stable (fluctuation ≤0.5 points), mild decline (decline 0.5-1 points), rapid degradation (decline >1 point) ”;

[0217] The model prediction accuracy is greater than or equal to 85%.

[0218] The embodiment has the following beneficial effects:

[0219] Overall evaluation comprehensiveness: integrating “local capability indicators + individual background information” avoids the one-sidedness of a single indicator (for example, “reaction speed 5 points” may be normal for an 80-year-old and abnormal for a 60-year-old), and the evaluation result is more in line with the actual situation of the individual;

[0220] Nonlinear relationship modeling accuracy: using MLP / GNN / LightGBM models, the complex correlation between “age + disease + capability index” (for example, the superimposed effect of “diabetes + living alone” on the living ability) is effectively captured, and the prediction accuracy is greater than or equal to 85%, which is 10%-15% higher than that of a linear model;

[0221] Risk early warning timeliness: not only output the current living ability level, but also achieve forward-looking early warning through the risk index and the trend label (for example, the “rapid degradation” label triggers care intervention), avoiding passive response after the ability deteriorates;

[0222] Grade division practicality: the division of 0-5 levels conforms to the industry standard (reference Barthel index, Lawton IADL scale), which is convenient for caregivers to quickly understand (for example, level 3 corresponds to “partial dependence on care”), and the risk index provides a quantitative basis for individualized intervention (for example, risk index >0.7 suggests increasing the frequency of care).

[0223] In some embodiments, the prediction result of the second layer model is fed back to the first layer model through a feature feedback mechanism, specifically including:

[0224] Based on the living ability level and the capability degradation risk output by the second layer model, SHAP value analysis is used to determine the behavior and cognitive ability dimensions that have a significant impact on the evaluation result (for example, in the “rapid degradation” risk, the SHAP absolute value of the reaction speed dimension is greater than or equal to 0.3);

[0225] According to the significant dimension, the weight coefficient of the corresponding feature in the first layer model is adjusted (such as the reaction speed related feature weight is increased by 20%-50%), and the sensitivity of the first layer model to the significant dimension change is enhanced (such as the reaction speed is decreased by 0.5 points to trigger a risk warning).

[0226] The feedback adjustment is repeated every 7-30 days based on new evaluation data, and the first layer model parameters are updated using an incremental learning strategy, so that the evaluation accuracy of the model remains ≥85% in long-term operation, and the generalization ability is improved by 10%-15%.

[0227] The present application improves the sensitivity and long-term stability of the first layer model to key ability changes through "key dimension identification, weight adjustment, and incremental update", as follows:

[0228] Key ability dimension identification:

[0229] Based on the life ability level and ability degradation risk output by the second layer model, SHAP value analysis is used to determine the behavior and cognitive ability dimensions that significantly affect the evaluation results - the absolute value of each ability dimension is calculated, and if the absolute value ≥0.3 (such as the SHAP absolute value of reaction speed in "rapid degradation" risk = 0.4), it is determined that this dimension is a key influencing dimension.

[0230] First layer model weight adjustment:

[0231] According to the key dimension, the weight coefficient of the corresponding feature in the first layer model is adjusted: if the reaction speed is the key dimension, the weight of the reaction speed related feature (such as the voice response time related feature) is increased from the base value (such as 0.2) by 20%-50% (to 0.24-0.3), and the sensitivity of the model to the reaction speed change is enhanced (such as the reaction speed is decreased by 0.5 points to trigger a risk warning).

[0232] Incremental update optimization:

[0233] Based on the newly collected evaluation data (such as 1000 samples) every 7-30 days, the first layer model parameters are updated using an incremental learning strategy - only the model weights are adjusted based on new samples, and the full data is not retrained; the knowledge of the old model is transferred to the updated model through knowledge distillation technology to ensure that the model performance does not decay; the evaluation accuracy of the updated model remains ≥85%, and the generalization ability is improved by 10%-15%.

[0234] This embodiment has the following beneficial effects:

[0235] Model sensitivity is improved: the feature weight of the key dimension is adjusted specifically, so that the core ability change (such as a slight decrease in reaction speed) can be identified in time, and key risks are avoided.

[0236] Long-term stability: Regular incremental updates ensure that the model adapts to changes in data distribution (e.g., slow degradation of elderly abilities), with evaluation accuracy remaining ≥85% over a 6-month operating period without significant decay;

[0237] Enhanced generalization ability: Feedback mechanism enables the model to stably evaluate in different individuals and scenarios (e.g., generalization error ≤5% in home / community scenarios), reducing scenario adaptation costs;

[0238] Efficient closed-loop optimization: No need to retrain the full data, only incremental updates are needed to achieve optimization, reducing computational resource consumption (update time is shortened by 60% compared to full training).

[0239] The second aspect of the present application provides an independent living ability assessment system for the elderly, comprising:

[0240] A data acquisition unit is configured to acquire multi-modal raw data of a target elderly person, wherein the multi-modal raw data includes at least two of physiological data, behavior data, and interaction data;

[0241] A data processing unit includes a modality receiving module, a modality encoding module, a query interaction module, a feature compression module, and a mapping output module; the modality receiving module is configured to align multi-channel data timing, the modality encoding module is configured to independently encode each modality data, the query interaction module is configured to cross-modality fusion and missing data completion, the feature compression module is configured to reduce feature dimension, and the mapping output module is configured to generate standardized input data;

[0242] A data analysis unit includes a first layer model module, a second layer model module, and a feature feedback module; the first layer model module is configured to evaluate the behavior and cognitive ability indicators of the target elderly person, the second layer model module is configured to combine individual background information and the behavior and cognitive ability indicators to predict the living ability grade and ability degradation risk; the feature feedback module is configured to feed back the prediction results of the second layer model module to the first layer model module to adjust the feature weight and evaluation strategy;

[0243] A visual output unit is configured to output structured evaluation results and perform visual display, wherein the structured evaluation results include the living ability grade, the ability degradation risk, and the behavior and cognitive ability indicators; the visual forms include ability radar chart (dimension score distribution), risk trend line chart (risk change in the past 1-3 months), numerical labels (ability grade, risk index), and support mobile APP and desktop platform access.

[0244] The present application realizes the whole process evaluation of the independent living ability of the elderly through the cooperative work of the four modules of "data acquisition unit-data processing unit-data analysis unit-visualization output unit", as follows:

[0245] Data acquisition unit:

[0246] Integrated video equipment, virtual reality equipment, environmental sensors (door magnetic, pressure pad), wearable devices (smart bracelet), smart home terminal (smart mattress, voice assistant), collect multi-modal raw data according to "real-time / periodic / wake-up" strategy; The time stamp of the associated data and the unique identifier of the elderly are stored in the InfluxDB time series database, supporting data tracing and calling;

[0247] Data processing unit:

[0248] 5 functional modules:

[0249] Modality receiving module: adopt DTW algorithm to realize multi-device data time sequence alignment, deviation ≤0.1 second;

[0250] Modality encoding module: through ResNet50+LSTM, Transformer, FFT+statistical quantity calculation, complete independent encoding of each modality data;

[0251] Query interaction module: based on pre-training query vector and multi-head self-attention mechanism, realize cross-modal fusion and missing information completion;

[0252] Feature compression module: through AutoEncoder / PCA algorithm dimension reduction, generate 64-128 dimensional standardized feature vector;

[0253] Mapping output module: generate "[identifier, timestamp, feature vector, modality label]" structured format, transmit to data analysis unit;

[0254] Data analysis unit:

[0255] 3 functional modules:

[0256] First layer model module: integrate GRU time sequence module and light weight model, output behavior and cognitive ability indexes such as reaction speed and execution stability (error ≤0.5 points);

[0257] Second layer model module: adopt MLP / GNN / LightGBM, combine individual background information, output life ability level (0-5 level) and degradation risk (risk index + trend label);

[0258] Feature feedback module: through SHAP value analysis and incremental learning, the second layer result is fed back to the first layer model to optimize the feature weight;

[0259] Visual output unit:

[0260] The evaluation result of the received data analysis unit is displayed through a capability radar chart (dimension score distribution), a risk trend line chart (change in the last 1-3 months), and a numerical label card (color-coded risk). The data transmission is encrypted using HTTPS. The system supports mobile APP (for family members to view), community care platform (for caregivers to develop plans), and hospital system (for doctors to refer) access.

[0261] This embodiment has the following beneficial effects:

[0262] Modular design is clear: the functions of each unit are independent and the interfaces are unified, facilitating maintenance and upgrading (e.g., replacing sensors only requires adjusting the data acquisition unit without affecting other modules);

[0263] High degree of automation: from data collection to result output, no human intervention is required, and 288 evaluations can be completed daily (5 minutes per evaluation), significantly reducing the workload of caregivers;

[0264] Multi-scenario adaptability: can be flexibly deployed in home (non-invasive devices are the main ones), community (public area cameras are supplementary), and hospital (combined with medical records) scenarios without the need for customized modifications;

[0265] Easy to understand and apply results: visual forms reduce the understanding threshold for non-professionals (family members can quickly identify risk levels), and structured output can directly interface with the elderly service process (e.g., communities can adjust visit frequency based on ability level).

[0266] In a third aspect, the present application provides a computer device, which refers to Figure 5 The computer device includes a memory and a processor, the memory stores a computer program, and the processor is configured to acquire the computer program and execute the above-mentioned method for evaluating the independent living ability of the elderly.

[0267] In some embodiments, the method for evaluating the independent living ability of the elderly in the above-mentioned embodiments can be implemented by a computer device, which includes at least one processor, a communication bus, a memory, and at least one communication interface.

[0268] The processor can be a general central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0269] A communication bus can be used to communicate information between the aforementioned components.

[0270] The memory can be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM), or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disk storage, a magnetic disk storage or other magnetic storage devices, or any other medium capable of storing desired program code in the form of instructions or data structures and that can be accessed by a computer, but not limited to. The memory can exist independently, and is connected with the processor through the communication bus. The memory can also be integrated with the processor.

[0271] The memory is configured to store program code for implementing the scheme of the present application, and the processor is configured to execute the program code. The processor is configured to execute the program code stored in the memory. The program code can include one or more software modules. The method for evaluating the independent living ability of the elderly in the above embodiments can be implemented by one or more software modules in the program code of the processor and the memory.

[0272] The communication interface is configured to communicate with other devices or communication networks using any transceiver-like device, such as an Ethernet, a radio access network (RAN), a wireless local area network (WLAN), etc.

[0273] In a specific implementation, as an embodiment, the computer device can include a plurality of processors, each of which can be a single-CPU processor or a multi-CPU processor. The processor here can refer to one or more devices, circuits, and / or processing cores for processing data (e.g., computer program instructions).

[0274] The computer device described above can be a general-purpose computer device or a special-purpose computer device. In a specific implementation, the computer device can be a desktop computer, a laptop computer, a network server, a personal digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. The embodiments of the present application do not limit the type of computer device.

[0275] In addition, the present application also provides a computer-readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the above-mentioned method for evaluating the independent living ability of the elderly. The computer program for implementing the above-mentioned method is stored in the computer-readable storage medium, so that the point distribution method has the characteristics of "storable, portable, and reusable". On the one hand, the storage medium can solidify the method into an executable program, avoiding the need to redevelop the algorithm every time the point distribution optimization is performed, thereby reducing the technical implementation cost. On the other hand, the program can be deployed on different computer devices, facilitating the promotion and application of the program in different regions and different scale power grid projects, and accelerating the engineering landing and industrialization promotion of the "evaluation technology for the independent living ability of the elderly".

[0276] In order to more clearly understand the present application, the present application is further described below:

[0277] An evaluation system for the independent living ability of the elderly specifically includes a data acquisition unit 101, a data processing unit 102, a data analysis unit 103, and a visual output unit 104, and the structure is shown in FIG. 1. The functions of each of the above components are introduced below.

[0278] The data acquisition unit 101 is used to acquire the original physiological data and original activity data of the target monitoring object. The data can be acquired by a video device, a virtual reality device, a sensor, and a wearable device. The acquisition content includes, but is not limited to, heart rate, blood oxygen, gait, action posture, voice state, and interactive behavior. The sensor device includes an acceleration sensor, a gyroscope, a bioelectric sensor, a temperature and humidity sensor, etc. The wearable device can include a smart bracelet, a wearable camera, etc. The system supports two modes of periodic acquisition and real-time acquisition, and performs standardized transmission through a unified data interface protocol.

[0279] The above-mentioned acquisition data is only an example given by the present application and is not used to limit the present application. In actual application, other historical data acquisition can be added to acquire the historical work and rest data and medical record data of the monitoring object in the window period, and to fuse with the current data, thereby providing a longitudinal reference basis for subsequent analysis.

[0280] The data processing unit 102 is configured to uniformly model, learn missing data and compress the raw physiological data and the raw activity data acquired by the data acquisition unit 101, to obtain a data feature vector.

[0281] Specifically, the data processing unit 102 receives the raw multi-modal data collected by the data acquisition unit 101, and performs format standardization, modal alignment, feature extraction and compression processing on the data, to generate a data feature vector with uniform structure, thereby providing high-quality input support for subsequent analysis layer modeling. The data processing layer module also has data cleaning and missing data completion capability, which supports robust modeling in the case of incomplete multi-modal data, and improves the overall stability and usability of the system.

[0282] The data analysis unit 103 is configured to input the data feature vector into a trained deep learning model, to perform multi-dimensional assessment and prediction of the independent living ability of the elderly, and to generate assessment result data.

[0283] Specifically, the data analysis unit 103 receives the target data output by the data processing unit 102, and performs analysis by using a hierarchical modeling architecture. The hierarchical model includes: a first layer model for modeling behavior and cognitive ability related to behavior execution, reaction speed and attention, and outputting behavior and cognitive ability assessment including but not limited to reaction speed, execution stability and attention concentration; and a second layer model for predicting the overall living ability grade or degradation risk trend of the elderly based on the ability assessment output by the first layer, and combining individual background information and historical data.

[0284] The visualization output unit 104 is configured to perform visualization processing on the assessment result, and present the assessment result and intervention suggestions of the living ability of the elderly to the user through an interactive interface.

[0285] Specifically, the visualization output unit 104 receives the assessment result data from the data analysis layer module, and displays the assessment content in a graphical manner by using various visualization methods, including ability score graph, risk trend graph and life function radar chart, to improve the understanding of the user on the result.

[0286] As shown in FIG. 2, the data processing unit 102 provided in the embodiment includes a modal receiving unit 1021, a modal encoding unit 1022, a query interaction unit 1023, a feature compression unit 1024 and a mapping output unit 1025.

[0287] The modal receiving unit 1021 is configured to receive the raw multi-modal data from the data acquisition unit 101. The modal receiving unit supports multi-channel synchronous acquisition and timestamp recording, to ensure the consistency and timing alignment of the multi-modal data, thereby providing a raw input basis for subsequent processing.

[0288] The modal encoding unit 1022 is used for independent encoding of various types of raw data obtained by the modal receiving unit 1021, converting them into representation vectors in a unified feature space, and forming second target (encoded) data.

[0289] Specifically, visual modal data can extract static and dynamic visual features through an image encoding network; action modal data can extract behavior features through a time series modeling method; and physiological modal data can generate a physiological state vector through frequency domain analysis, statistical quantity extraction, and other methods. After encoding, the features of each modality are projected into a unified vector space, facilitating subsequent cross-modal fusion and semantic alignment.

[0290] The query interaction unit 1023 is used to construct a set of learnable query vectors, interact with the multi-modal features output by the modal encoding unit, realize data alignment and key feature extraction, and output third target data.

[0291] Specifically, the query interaction unit 1023 introduces a set of parameterized query vectors as a task-oriented feature extractor. In this process, the system automatically selects the most representative data features for the current evaluation task through a multi-modal pre-training task mechanism, realizing data alignment and fusion between different modalities. The query interaction unit 1023 has modality independence and feature selectivity, and can dynamically select effective feature sources and extract cross-modal shared information in the case of incomplete or uneven quality of modal data, thereby improving the consistency and representativeness of overall feature expression. The third target data output by the query interaction unit 1023, as the fused representation, has good generality and interpretability.

[0292] The feature compression unit 1024 is used for compression processing of the fused features output by the query interaction unit, generating a compact and representative unified representation vector, and outputting fourth target data.

[0293] Specifically, the feature compression unit 1024 uses dimension reduction methods (such as AutoEncoder, MLP) to compress the fused feature vector, retaining information essential for life ability evaluation and removing redundant dimensions. This representation vector has the advantages of low dimensionality, high expressiveness, and good interpretability, facilitating fast processing by subsequent deep analysis models.

[0294] The mapping output unit 1025 is used to convert the representation vector output by the feature compression unit into a structured target data format conforming to the system standard, as the input of the data analysis layer module.

[0295] Specifically, the mapping output unit 1025 performs format normalization, field labeling and meta information binding operations on the feature vector, and outputs content including capability-related feature vectors, acquisition timestamps, object identifiers, modal labels and the like, forming final target data. The target data structure is clear and interface-friendly, and can be seamlessly connected with subsequent modules of the system.

[0296] In view of the difficulty of the existing system in multi-source heterogeneous data fusion, and the weak processing ability for data missing, it is difficult to maintain analysis accuracy and stability in the case of incomplete information. The present application discloses a multi-modal pre-training mechanism, which comprises:

[0297] Modality alignment task: requires the model to model the data consistency of different modalities (such as image and action, action and physiology, etc.), that is, to learn to map the representations of different modalities to the same vector space;

[0298] Modality matching discrimination task: for the input multi-modal combination (including matching samples and non-matching samples), the model is trained to judge whether the different modal data belong to the same person, which is used to enhance the ability of the model to identify data differences;

[0299] Behavioral semantic modeling task: based on the fusion features, a structured capability description or behavior summary is generated, which is used to improve the ability of the model to understand data and interpretability.

[0300] Through the above pre-training tasks, the query interaction unit 1023 can not only extract key capability-related features from complex data, but also achieve efficient data fusion and task generalization.

[0301] As shown in Figure 3 The data analysis unit 103 provided by the embodiment includes a first layer model unit 1031, a second layer model unit 1032 and a feature feedback mechanism unit 1033, which are used to model and predict the independent living ability of the elderly based on the input feature vector, and output structured evaluation results.

[0302] The first layer model unit 1031 is used to model the ability state of the elderly in behavior and cognition based on the compressed feature vector output by the data processing module, and outputs intermediate capability indicators including but not limited to reaction speed, execution stability, attention concentration, etc.

[0303] The unit can integrate a time series modeling module to extract time-related features, and combine a lightweight modeling structure (such as logistic regression or regression network) to perform multi-dimensional capability evaluation, with strong interpretability and real-time performance.

[0304] The output intermediate capability vector is used as the input of the second layer model, and can also be used for independent visual analysis and individual capability profile display.

[0305] The second layer model unit 1032 is used to complete the life ability level prediction and degradation risk trend analysis based on the first layer intermediate ability indicators, combined with individual background information (such as age, health records, historical behavior data, etc.).

[0306] The model can use a nonlinear structure (such as a multilayer perception machine, a graph neural network, etc.) to deeply model high-order features, learn the complex relationship between the ability change trend and the potential risk factors. The output results include: life ability level (such as 0-5 levels), ability dimension score (such as communication, processing, movement, cognition, etc.), ability trend label (such as stable, mild decline, rapid degradation), and risk index (such as: degradation risk score 0.84), with individualized prediction ability (such as: suggest increasing exercise every week).

[0307] The feature feedback mechanism unit 1033 is used to feed back the analysis results of the second layer model to the first layer model, to realize feature weight adjustment and strategy optimization. This mechanism dynamically corrects the feature selection or weight distribution of the first layer model based on the evaluation output results, improves the sensitivity and robustness of the system to the changes of key ability dimensions, and thus enhances the stability, generalization ability and interpretability of the overall model.

[0308] Technical effect summary

[0309] The elderly ability assessment system provided by the present application aims to solve the problems of low assessment frequency, strong subjectivity and poor real-time performance caused by relying on artificial questionnaires, on-site observation or single physiological indicator judgment in the existing assessment methods. The system realizes continuous monitoring and quantitative analysis of the behavior, cognition and physiological state of the elderly through intelligent perception and fusion modeling of multi-modal data. The objectivity, frequency and individual matching degree of the assessment are significantly improved, and the system is suitable for various application scenarios such as home care, community monitoring and smart medical care.

[0310] The data processing unit 102 provided in the embodiment can efficiently complete the processing procedure from the original multi-modal data to the standardized representation data. Through the processing steps of modality receiving, encoding, semantic interaction, compression and mapping, the effective fusion and feature extraction of multi-modal information are realized, and high-quality input is provided for the subsequent life ability assessment model. Even if there is modal missing or data noise, the system can still stably output key features, and improve the robustness and generalization of the elderly ability assessment model.

[0311] The data processing module 103 provided by the embodiment adopts a hierarchical modeling structure. The first layer model evaluates cognitive and behavioral ability indexes such as reaction speed, execution stability, and attention concentration. The second layer model combines individual background information to perform life ability grade evaluation and degradation risk prediction, thereby improving individualized analysis effect. Meanwhile, through a feature feedback mechanism, the second layer model can reversely optimize the feature strategy and weight distribution of the first layer model, thereby enhancing the stability and explanation ability of the overall model.

[0312] To further illustrate the feasibility of the present application, the following specific implementation cases are given in combination with the home-based elderly care scene:

[0313] I. Implementation environment

[0314] Target object: 65-year-old elderly A living alone (with a history of hypertension, using an intelligent bracelet and a home camera daily);

[0315] Device deployment:

[0316] The intelligent bracelet (Huawei Watch GT5) is worn on the left wrist of the elderly, and collects heart rate, blood oxygen, and motion;

[0317] The door magnetic sensor (Xiaomi intelligent door magnetic) is installed on the doors of the bedroom and the kitchen, and records the opening and closing times;

[0318] The panoramic camera (Hikvision DS-2CD3T46DWD-I3) is installed in the living room to collect daily motion;

[0319] The voice assistant (Xiaoyi) is placed in the living room to record voice interaction;

[0320] Data storage and calculation: The edge computing node (Raspberry Pi 4B) deploys the InfluxDB database and the model inference environment, and the data is transmitted to the node in real time, and the model inference time is ≤1 second / second.

[0321] II. Implementation process

[0322] Data collection: May 1-7, 2025, continuously collect multi-modal data of the elderly A, among which 4 hours of gait data are missing on May 3 due to power failure of the intelligent bracelet;

[0323] Data processing:

[0324] Modality encoding: The camera data outputs 512-dimensional gait features, the intelligent bracelet (after power recovery) data outputs 256-dimensional motion features, and the heart rate data outputs 128-dimensional physiological features;

[0325] Query interaction: For the missing gait data on May 3rd, the model complements the gait features through "heart rate fluctuation (smooth, no abnormalities) + door magnetic record (kitchen door opening and closing 2 times, reflecting autonomous activities)", with a complementation error ≤10%;

[0326] Feature compression: Output 128-dimensional normalized feature vector;

[0327] Hierarchical analysis:

[0328] First layer model output: Reaction speed 8.5 points (average response time to voice instructions 0.8 seconds), execution stability 7.2 points (standard deviation of standing pressure 85kPa), attention concentration 9.0 points (device operation without error), motion coordination 8.0 points (step length variance 8cm);

[0329] Second layer model combines the age of the old man A (65 years old) and the history of hypertension, output: life ability level 5 (completely independent), risk index 0.12 (low risk), risk trend label "stable", key influencing factor "good motion coordination, normal reaction speed";

[0330] Feature feedback: Because the risk trend is stable, the first layer model weight is not adjusted;

[0331] Visual output: The family members of the old man A view the ability radar chart (each dimension score ≥7 points) and the risk trend chart (the risk index is stable at 0.1-0.15 in the past 1 week) through the APP, and the system gives the intervention suggestion "suggest maintaining 30 minutes of daily walking to maintain motion coordination".

[0332] III. Implementation effect verification

[0333] Evaluation accuracy: Consistent with the artificial evaluation results of medical staff at the same period (life ability level 5, low risk);

[0334] Robustness of missing data: When the gait data is missing on May 3rd, the evaluation accuracy only decreases by 2% (from 88% to 86%), which is much better than the existing technology (decreases by more than 30%);

[0335] Interpretability: Clearly points out that "the core reason for stable ability is good reaction speed and motion coordination", which is convenient for family members to provide targeted care;

[0336] Practicality: Community caregivers adjust the frequency of home visits for the old man A from 2 times a week to 1 time a week based on the output "low risk" conclusion, optimizing resource allocation.

[0337] The technical solutions provided by the embodiments of the present application are described in detail above, and the principles and implementation manners of the embodiments of the present application are described by applying specific examples. The above description of the embodiments is only used to help understand the principles of the embodiments of the present application; meanwhile, for those skilled in the art, the embodiments of the present application will have changes in specific implementation manners and application scopes, and the above description should not be understood as a limitation on the present application.

Claims

1. A method for assessing the independent living ability of the elderly, characterized in that, Includes the following steps: Collect multimodal raw data of the target elderly, wherein the multimodal raw data includes at least two of the following: physiological data, behavioral data, and interaction data; The multimodal raw data is processed, including modality encoding, query vector interaction fusion and missing information completion, and feature compression, to obtain standardized feature vectors; Based on the standardized feature vector, a hierarchical analysis is performed. The first-layer model evaluates the behavioral and cognitive ability indicators of the target elderly person. The second-layer model combines individual background information and the behavioral and cognitive ability indicators to predict the level of living ability and the risk of ability decline. The prediction results of the second-layer model are fed back to the first-layer model through a feature feedback mechanism, and the feature weights and evaluation strategies of the first-layer model are adjusted. The system outputs the structured assessment results of the life skills level, the risk of skill degradation, and the behavioral and cognitive ability indicators, and presents them visually.

2. The method for assessing the independent living ability of the elderly according to claim 1, characterized in that: The multimodal raw data collected from the target elderly population specifically includes: The multimodal raw data is collected by at least two sensing devices, including at least two of wearable devices, environmental sensors, video devices, virtual reality devices, and smart home terminals. The physiological data includes at least one of heart rate, blood oxygen saturation, blood pressure, and respiratory rate; the behavioral data includes at least one of gait characteristics, movement posture, and daily activity trajectory; and the interaction data includes at least one of voice interaction content and device operation records. It supports real-time and periodic data collection, and associates the timestamps of the multimodal raw data with the unique identifier of the target elderly person.

3. The method for assessing the independent living ability of the elderly according to claim 1, characterized in that: The processing of the multimodal raw data specifically includes: The multimodal raw data is received synchronously through the modal receiving unit, and the timing consistency calibration of multi-channel data is achieved by using a timestamp alignment algorithm; The original data of each modality are encoded separately through the modality coding unit: the visual modality data is processed by a convolutional neural network + temporal network to extract dynamic behavior features, the action modality data is processed by a Transformer encoder to extract behavior sequence features, and the physiological modality data is processed by a fast Fourier transform + statistical calculation to extract physiological state features, thus obtaining the representation vector of each modality; Learnable query vectors are constructed through query interaction units, and attention mechanisms are used to interact with each of the representation vectors to achieve cross-modal feature alignment and extraction of key capability-related features. The dimensionality reduction of the interactively fused features is performed by the feature compression unit to obtain the standardized feature vector with dimensions of 64-128. The standardized feature vector is formatted by the mapping output unit, and the data collection timestamp, target elderly identifier, and modality validity label are bound to generate input data for hierarchical analysis.

4. The method for assessing the independent living ability of the elderly according to claim 3, characterized in that: Before the query interaction unit constructs a learnable query vector for interactive operation, it also includes: The query vector is pre-trained in a multimodal manner, and the pre-training task includes a combination of modality alignment task, modality matching and discrimination task, and behavior semantic modeling task. The modality alignment task employs a contrastive learning strategy to map the representation vectors of different modalities to the same feature space, ensuring that the cosine similarity of the multimodal vectors corresponding to the same behavior is ≥0.

8. The modality matching and discrimination task constructs a binary classification model, which is trained to identify whether the input multimodal data belongs to the same target elderly person, with an accuracy of ≥90%. The behavioral semantic modeling task adopts the Seq2Seq model to generate a structured description of the behavioral abilities of the target elderly based on fused features.

5. The method for assessing the independent living ability of the elderly according to claim 3, characterized in that: The feature compression unit performs dimensionality reduction on the interactively fused features, specifically employing at least one of AutoEncoder, multilayer perceptron, principal component analysis, t-distributed random neighborhood embedding, or sparse representation methods, to retain key features related to the assessment of living abilities and remove redundant dimensions.

6. The method for assessing the independent living ability of the elderly according to claim 1, characterized in that: The first-layer model assesses the behavioral and cognitive indicators of the target older adults, specifically including: The first-layer model integrates a time-series modeling module to extract the time-related features of the standardized feature vector within a continuous acquisition period; The behavioral and cognitive ability indicators are output using at least one of logistic regression, lightweight regression network, or decision tree models. The behavioral and cognitive ability indicators include reaction speed score, executive stability score, attention concentration score, and motor coordination score, with a calculation error of ≤0.5 points for each indicator.

7. The method for assessing the independent living ability of the elderly according to claim 1, characterized in that: The second-layer model combines individual background information with the aforementioned behavioral and cognitive ability indicators to predict the level of daily living skills and the risk of skill decline, specifically including: The individual background information includes age, history of underlying diseases, historical ability assessment records, and living environment characteristics, which are transformed into vector features through an embedding layer. The behavioral and cognitive ability indicators are concatenated with the individual background information vector to obtain the input features; The second-layer model employs at least one of a multilayer perceptron, a graph neural network, or a gradient boosting tree model to learn the nonlinear correlation between the behavioral and cognitive ability indicators and living abilities. The output of life skills level is divided into 0-5 levels, and the risk of skill degradation includes risk index and risk trend label.

8. The method for assessing the independent living ability of the elderly according to claim 1, characterized in that: The step of feeding back the prediction results of the second-layer model to the first-layer model through a feature feedback mechanism specifically includes: Based on the living ability level and ability degradation risk output by the second-layer model, SHAP value analysis was used to determine the behavioral and cognitive ability dimensions that significantly affect the assessment results; Adjust the weight coefficients of the corresponding features in the first layer model according to the salient dimension to enhance the sensitivity of the first layer model to changes in the salient dimension; The first-layer model parameters are updated every 7-30 days based on new evaluation data, using an incremental learning strategy.

9. A system for assessing the independent living ability of the elderly, characterized in that, include: The data acquisition unit is used to collect multimodal raw data of the target elderly person, wherein the multimodal raw data includes at least two of the following: physiological data, behavioral data, and interaction data. The data processing unit includes a modality receiving module, a modality encoding module, a query interaction module, a feature compression module, and a mapping output module. The modality receiving module is used for time-series alignment of multi-channel data, the modality encoding module is used for independent encoding of each modality data, the query interaction module is used for cross-modality fusion and missing data completion, the feature compression module is used for feature dimensionality reduction, and the mapping output module is used to generate standardized input data. The data analysis unit includes a first-layer model module, a second-layer model module, and a feature feedback module; The first-layer model module is used to evaluate the behavioral and cognitive ability indicators of the target elderly person; the second-layer model module is used to combine individual background information and the behavioral and cognitive ability indicators to predict the level of living ability and the risk of ability decline; the feature feedback module is used to feed back the prediction results of the second-layer model module to the first-layer model module to adjust the feature weights and evaluation strategies. The visualization output unit is used to output and visualize the structured assessment results, which include the living ability level, ability decline risk, and behavioral and cognitive ability indicators. The visualization forms include ability radar charts, risk trend line charts, and numerical labels, and support access from mobile apps and desktop platforms.

10. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method for assessing the independent living ability of the elderly as described in any one of claims 1-8.

Images