An old person internal ability multi-mode dynamic evaluation method and cloud platform

Abstract

The application belongs to the technical field of medical data analysis, and discloses a kind of old person internal ability multimodal dynamic evaluation method and cloud platform;Including, receive multi-source health data on cloud platform, form multimodal original data set organized according to evaluation time point, deploy acute event independent monitoring and disposal mechanism;Calculate the single-domain Z score of five fields;Based on the single-domain Z score of five fields and multimodal original data set, construct the internal ability multimodal time series feature vector of the old person to be tested at multiple evaluation time points;Deploy internal ability dynamic evaluation model on cloud platform, output the internal ability individualized comprehensive score of current evaluation time point, and the future change trend index in future preset time window;According to individualized comprehensive score and future change trend index, determine the current internal ability risk level of old people, and adaptively adjust evaluation strategy and data acquisition mode;Realize the multimodal dynamic evaluation of old people's internal ability.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and more specifically, to a multimodal dynamic assessment method and cloud platform for the intrinsic abilities of the elderly. Background Technology

[0002] Traditional elderly health management models often focus on the diagnosis and treatment of single diseases or the prevention of specific functional impairments (such as frailty and disability), lacking a comprehensive and dynamic assessment of an individual's overall physiological and psychological reserves. Intrinsic capacity is defined as the sum of all an individual's physical and mental capabilities, encompassing five key areas: movement, vitality (nutrition), cognition, psychology, and sensory perception. As a dynamic indicator, intrinsic capacity emphasizes the importance of continuous monitoring and assessment throughout the life cycle, providing a new theoretical framework and assessment targets for early risk identification and preventative intervention.

[0003] At the intervention level, given the complexity of health problems among the elderly, involving multiple coexisting diseases and intertwined risks, single-type interventions (such as simple exercise or nutritional supplementation) often have limited effectiveness. In recent years, the concept of "multi-domain intervention" has gradually emerged. Its core lies in integrating two or more evidence-based intervention strategies to produce a synergistic effect and more comprehensively improve the overall functional status of the elderly. Among them, comprehensive intervention programs combining exercise training, cognitive training, and nutritional guidance have been confirmed by multiple studies to have positive effects on improving common geriatric syndromes such as frailty, sarcopenia, cognitive impairment, and depression.

[0004] However, existing technologies face the following major problems and challenges in practice:

[0005] Although the concept of multi-domain intervention has been accepted, in practice, interventions such as exercise, cognition, and nutrition are often provided separately by personnel from different professions, resulting in a fragmented approach lacking systematic integration and coordination. This makes it difficult to achieve a synergistic effect where "1+1>2," leading to fragmented assessments of older adults' abilities. There is a lack of a unified, standardized framework to integrate and quantify the five major domains of exercise, nutrition, cognition, psychology, and sensory functions. Assessments are often one-off or long-term, failing to achieve dynamic and continuous monitoring of intrinsic abilities, making it difficult to capture subtle changes and adjust intervention strategies in a timely manner.

[0006] In view of this, a multimodal dynamic assessment method and cloud platform for the intrinsic abilities of the elderly are designed. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a multimodal dynamic assessment method for the intrinsic abilities of the elderly, comprising:

[0008] S1. Receive multi-source health data on the cloud platform, perform time alignment, form a multimodal raw dataset organized according to the assessment time point, deploy an independent monitoring and handling mechanism for acute events, and calculate single-domain Z-scores for five domains by combining pre-stored data.

[0009] S2. Based on the single-domain Z-scores of five domains and the original multimodal dataset, construct the multimodal time series feature vector of the intrinsic abilities of the elderly under test at multiple assessment time points;

[0010] S3. Deploy the intrinsic capability dynamic assessment model on the cloud platform, input multimodal time series feature vectors, and then output the individualized comprehensive score of intrinsic capability at the current assessment time point, as well as the future change trend indicators within the future preset time window;

[0011] S4. Predefined intrinsic capability risk level;

[0012] S5. Based on individualized comprehensive scores and future trend indicators, determine the current intrinsic risk level of the elderly and adaptively adjust the assessment strategy and data collection mode.

[0013] Preferably, the multimodal raw data is used to represent the data that needs to be collected from the elderly in five domains: cognition, movement, sensory, nutrition and psychology. It includes core scale data and real-time data. The real-time data includes behavioral data, physiological data, activity data, strength data, dietary data, metabolic data and social and behavioral data.

[0014] The core scale data includes the raw scale scores for each domain. The cognitive domain is assessed directly using the Mini-Mental State Scale, the motor domain is assessed directly using the Mini-Mental State Scale, the sensory domain is assessed directly using the Visual and Auditory Self-Report, the nutritional domain is assessed directly using the Mini-Nutrition Assessment, and the psychological domain is assessed directly using the Geriatric Depression Scale.

[0015] The generation cycle of the nearest assessment time point T_n is dynamically set according to the intrinsic ability risk level. For core scale data, the effective assessment result closest to the assessment time point T_n is used. For real-time data, a dynamic time window is defined that goes back from T_n. The length of the window is linked to the intrinsic ability risk level. Within this window, features are extracted from the real-time data. For each extracted feature, Z-score is used for standardization and normalization to the [0,1] interval.

[0016] Preferably, the method for deploying an independent monitoring and response mechanism for acute events includes:

[0017] Establish an individualized dynamic baseline based on the activity and physiological data of each elderly person;

[0018] Based on individualized dynamic baselines, dynamic anomaly detection thresholds are set for each indicator, including thresholds for sudden drops or rises in a single day and thresholds for continuous anomalies. Predefined judgment rules are used, and cross-validation logic is introduced. When any indicator reaches its dynamic anomaly detection threshold, a primary anomaly signal is generated. During the same time period when the primary anomaly signal is generated, the status of other indicators is checked. If at least two indicators trigger primary anomaly signals at the same time, a high-confidence acute anomaly indicator is immediately generated. If the anomaly of a single indicator continues for more than a preset duration, it is upgraded to an acute anomaly indicator.

[0019] When an acute abnormal indicator is triggered, an acute health risk warning is generated and sent to the preset guardian; the risk level is immediately switched to high risk, and an emergency assessment is arranged according to the first adjustment strategy.

[0020] Preferably, the method for calculating single-domain Z-scores across five domains by combining pre-stored data includes:

[0021] We collected data from the five core scales of a healthy older adult population, stratified them according to three key demographic variables: age, gender, and education level, and calculated the mean and standard deviation of the scores in each of the five domains within each stratum to form a reference population database.

[0022] When assessing a specific elderly person, the system automatically identifies the elderly person's age group, gender, and education level, and uses these as an index to retrieve and match the mean and standard deviation of the corresponding stratum from the reference population database, which are then recorded as pre-stored data.

[0023] For each domain, calculate the single-domain Z-score as: (Individual measurement - Mean of the reference population for the matched stratification) ÷ Standard deviation of the reference population for the matched stratification.

[0024] Preferably, the method for constructing the multimodal time-series feature vector of the intrinsic abilities of the elderly person to be tested at multiple assessment time points includes:

[0025] For a single assessment time point, the current assessment time point is used as the anchor point, and the single-domain Z-scores and core scale data of the five domains are directly incorporated; within the dynamic time window, all extracted features are concatenated to form a time-slice multimodal feature vector.

[0026] Each data point in the time-slice multimodal feature vector is assigned a dynamic confidence weight. For core scale data, if it is completed within the dynamic time window specified by T_n, the weight is 1.0. If it exceeds this period, the weight decays over time, but the minimum weight is retained. For real-time data, the weight is calculated based on the data completeness rate within the dynamic backtracking window. If the completeness rate is higher than the set completeness threshold, the weight is 1.0. If it is lower than the threshold, the weight is reduced proportionally.

[0027] All features in the same domain, after being weighted, are concatenated to form feature sub-vectors for each domain. These sub-vectors are then concatenated to form a complete time-slice multimodal feature vector. Subsequently, the time-slice multimodal feature vectors generated at each evaluation time point are arranged in chronological order according to their corresponding timestamps to obtain an intrinsic capability multimodal time series feature vector.

[0028] Preferably, the intrinsic capability dynamic assessment model includes, in sequence, a first level, a second level, and a third level:

[0029] The input of the first level is a multimodal time series feature vector, which includes an individualized baseline encoder and a multi-domain deep encoder. The input of the individualized baseline encoder is the first k time slice multimodal feature vectors in the multimodal time series feature vector. The structure includes domain baseline encoding and temporal aggregation.

[0030] The domain baseline encoding consists of five lightweight encoders with shared weights, each with 32 units and activated by the ReLU function. Each encoder processes five domain feature sub-vectors at each evaluation time point and outputs the domain baseline encoding. Temporal aggregation is used to input the domain baseline encodings of the G evaluation time points into a unidirectional LSTM in chronological order. The LSTM has 64 hidden units and passes through a 32-unit fully connected layer to generate individualized baseline vectors.

[0031] The input to the multi-domain deep encoder is the time-slice multimodal feature vector at the current evaluation time point T_n. It consists of five domain coding networks with the same structure but independent parameters working in parallel. Each network consists of two fully connected layers, set to 64 units and 32 units respectively, activated by the ReLU function, and the output is a deep embedding vector corresponding to each domain.

[0032] The second layer takes the output of the first layer as input, the individualized baseline vector as prior condition, and adds it to the time slice multimodal feature vector at each evaluation time point T_n. It uses three sets of one-dimensional dilated causal convolutions with dilation rates of 1, 2 and 4 and a kernel size of 3. Then it is input into a two-layer GRU network with 128 and 64 hidden units respectively. The final state of the last GRU layer is taken as the temporal context vector.

[0033] The third-level input consists of the time-slice multimodal feature vector and temporal context vector at the current evaluation time point T_n. Five domain feature vectors are used as queries, and the temporal context vector is used as the key. Four-head attention calculations are performed, each head with 16 dimensions. When calculating attention weights, the average confidence weight of each domain is added as a prior bias to the Softmax calculation. The resulting attention-weighted domain representation is concatenated with the temporal context vector and fused through a gating mechanism to output the final comprehensive state representation. Then, two fully connected layers are passed. One fully connected layer is set to 1 unit and uses a sigmoid function to output a normalized score, denoted as the individualized comprehensive score of intrinsic ability. The other fully connected layer is set to 2 units and uses a linear function to output two values: a predicted comprehensive score and a probability of decline risk, denoted as the future trend indicator.

[0034] Preferably, the training method for the intrinsic ability dynamic assessment model includes:

[0035] For each elderly person's data, training samples are constructed using a sliding window approach. Each sample contains an input sequence and supervision labels. The input sequence is a time-slice multimodal feature vector of length L with consecutive assessment time points. The supervision labels include current ability labels and future trend labels. The current ability label refers to the individualized comprehensive score marked by a domain expert at the last assessment time point of the sequence. The future trend label refers to the supervision signal of the actual comprehensive score change and whether a significant functional decline has occurred within a dynamic time window after the last assessment time point.

[0036] The total loss function is a time-consistency regularized weighted sum of the rating regression loss, trend prediction loss, and the rating regression loss. The rating regression loss is the smoothed L1 loss between the output of the individualized comprehensive rating model and the current ability label; the trend prediction loss is the mean squared error loss between the future trend indicator output by the model and the future trend label; and the time-consistency regularization is the L2 norm of the rating difference between adjacent time points.

[0037] The AdamW optimizer is used to iterate training in mini-batch mode. After each batch, the gradient is calculated and the main network parameters are updated through the optimizer. The individualized baseline encoder parameters are kept frozen.

[0038] Training is stopped when the total loss function value on the validation set no longer decreases for several consecutive rounds.

[0039] Deploy the trained model on a cloud platform.

[0040] Preferably, the predefined intrinsic capability risk level is jointly determined by an individualized comprehensive score and a future trend indicator; the future trend indicator includes at least the trend prediction slope and the trend prediction uncertainty.

[0041] The determination of a high-risk level requires the following conditions to be met simultaneously: the individualized comprehensive score is lower than the first static threshold, the trend prediction slope is negative, and the trend prediction uncertainty is lower than the preset uncertainty threshold.

[0042] The determination of a medium-risk level meets any of the following conditions:

[0043] Condition 1: The individualized comprehensive score is between the first static threshold and the second static threshold;

[0044] Condition 2: The individualized comprehensive score is higher than the second static threshold, but the trend prediction slope is negative and the trend prediction uncertainty is lower than the uncertainty threshold;

[0045] Condition 3: The uncertainty of trend prediction is higher than or equal to the uncertainty threshold;

[0046] The determination of a low-risk level requires the following conditions to be met simultaneously: the individualized comprehensive score is higher than the second static threshold, and the trend prediction slope is zero or positive.

[0047] Among them, the trend prediction slope refers to the average rate of change of the individualized comprehensive score of the intrinsic ability of the elderly within a preset time window in the future, and the uncertainty of change refers to the degree of grasp or confidence level of the model on the trend prediction slope.

[0048] Preferably, S5 includes:

[0049] The system pre-stores assessment period parameters corresponding to high-risk, medium-risk, and low-risk levels respectively; among them, the high-risk level is associated with the first period parameter, the medium-risk level is associated with the second period parameter, and the low-risk level is associated with the third period parameter, and the first period parameter < the second period parameter < the third period parameter.

[0050] When a high-risk level is determined, the first adjustment strategy is implemented: the period of the next assessment time point is set to the first period parameter, and the dynamic time window length is shortened simultaneously; within the window, the frequency of continuous data collection is increased, and the completeness threshold is reduced.

[0051] When the risk level is determined to be medium, the second adjustment strategy is implemented: the period of the next assessment time point is set to the second period parameter; based on the attention weight output of the third level of the intrinsic capability dynamic assessment model, the areas with attention weights higher or lower than the historical average are identified, and for the areas, the sampling density of relevant real-time data is increased in the subsequent dynamic time window.

[0052] When the risk level is determined to be low, the third adjustment strategy is implemented: the period for the next assessment time point is set to the third period parameter, while maintaining the existing data collection strategy.

[0053] A cloud platform for multimodal dynamic assessment of the intrinsic abilities of the elderly includes:

[0054] Data acquisition and acute monitoring module: used to receive multi-source health data, form a multimodal raw dataset organized according to the assessment time point, and deploy an independent monitoring and handling mechanism for acute events;

[0055] Multi-source data processing module: used to combine pre-stored data to calculate single-domain Z-scores in five domains, and to construct multi-modal time series feature vectors of intrinsic abilities of the elderly under test at multiple assessment time points based on the multimodal raw dataset;

[0056] Evaluation model building and training module: used to deploy dynamic evaluation models of intrinsic capabilities on cloud platforms;

[0057] Dynamic assessment module: used to obtain an individualized comprehensive score of intrinsic ability at the current assessment time point, as well as future trend indicators within a preset time window, based on the intrinsic ability dynamic assessment model;

[0058] Strategy generation module: Used to determine the current intrinsic ability risk level of the elderly based on individualized comprehensive scores and future trend indicators, and adaptively adjust the assessment strategy and data collection mode.

[0059] The technical effects and advantages of the multimodal dynamic assessment method for intrinsic abilities of the elderly in this invention are as follows:

[0060] By designing a multi-source heterogeneous data processing pipeline with "cloud collaboration and quality control integration", the problem of data with different sources, mixed frequencies, and uneven quality that are difficult to effectively integrate has been solved. It realizes the automated construction of high-quality, interpretable, and confidence-based standardized feature vectors from raw data, laying a reliable data foundation for accurate and intelligent evaluation.

[0061] By designing a proprietary multi-level deep learning model, the problem of general models being insufficient in capturing the slow, multi-dimensional, and individualized functional change patterns of the elderly is solved. This enables the model to create an individualized and accurate profile of the current state of internal capabilities, make scientific and forward-looking predictions of future decline trends, and endow the model with the "clinical thinking" ability to distinguish data quality and understand domain contributions.

[0062] By designing closed-loop management rules, the system addresses the issues of outdated risk assessment, rigid resource allocation, and insufficient acute response, achieving an adaptive health monitoring system that integrates "long-term, slow risk trend management" with "short-term, acute risk event response." This system acts like an experienced caregiver, providing appropriate attention and resources to elderly individuals in different risk states, achieving optimal efficiency while ensuring safety. Attached Figure Description

[0063] Figure 1This is a schematic diagram of the method for multimodal dynamic assessment of the intrinsic abilities of the elderly according to the present invention;

[0064] Figure 2 This is a schematic diagram of the multimodal dynamic assessment system for the intrinsic abilities of the elderly in this invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 and Figure 2 In this embodiment of the invention, a multimodal dynamic assessment method for the intrinsic abilities of the elderly includes:

[0067] S1. Receive multi-source health data from elderly terminals, wearable devices, and medical institution information systems on the cloud platform, perform time alignment, form a multimodal raw dataset organized according to the assessment time point, deploy an independent monitoring and handling mechanism for acute events, and calculate single-domain Z-scores in five domains by combining pre-stored data.

[0068] S2. Based on the single-domain Z-scores of five domains and the original multimodal dataset, construct the multimodal time series feature vector of the intrinsic abilities of the elderly under test at multiple assessment time points;

[0069] S3. Deploy the intrinsic capability dynamic assessment model on the cloud platform, input multimodal time series feature vectors, and then output the individualized comprehensive score of intrinsic capability at the current assessment time point, as well as the future change trend indicators within the future preset time window;

[0070] S4. Predefined intrinsic capability risk level;

[0071] S5. Based on individualized comprehensive scores and future trend indicators, determine the current intrinsic risk level of the elderly and adaptively adjust the assessment strategy and data collection mode.

[0072] The multimodal raw data is used to represent the data that needs to be collected from older adults in five domains: cognition, movement, sensory, nutrition and psychology. It includes core scale data and real-time data. Real-time data includes behavioral data, physiological data, activity data, strength data, dietary data, metabolic data and social and behavioral data.

[0073] Behavioral data in the cognitive domain refers to reaction time, accuracy, and trajectory in mobile cognitive training tasks; physiological data is used to represent sleep quality (deep sleep duration, sleep efficiency) and daytime activity regularity, providing evidence of physiological causes for cognitive fluctuations.

[0074] Activity data includes daily steps, walking speed, walking regularity, and standing time. Changes in walking speed are a highly sensitive indicator in the early stages of motor weakness, and are used to monitor continuous daily exercise capacity. Strength data refers to the trends in muscle strength and joint range of motion measured by wearable devices or smart instruments.

[0075] Behavioral data in the sensory domain refers to the frequency of using mobile phone font magnification, voice call volume adjustment records, and reaction time to audio and video stimuli. This data can realistically and imperceptibly reflect the actual use of the senses and compensatory behaviors.

[0076] Dietary data refers to near real-time monitoring of eating habits through images or text recording the types and portions of food consumed at three meals a day. Metabolic data refers to the continuous monitoring trends of weight and body composition (such as muscle mass and fat mass) in older adults, which can accurately assess the actual effect of nutritional interventions on muscle mass (rather than just weight). Therefore, both can be used to quantify data in the field of nutrition.

[0077] Social and behavioral data refers to the frequency of participation in group activities, the frequency of communication with family and friends, and mobile device usage activity, which are used to quantify objective behaviors related to mental health. Physiological data in the field of psychology also includes heart rate variability (HRV, which is an important physiological indicator reflecting emotional stress and autonomic nervous function) and resting heart rate.

[0078] The core scale data includes the raw scale scores for each domain. The cognitive domain is assessed directly using the Mini-Mental State Scale, the motor domain is assessed directly using the Mini-Mental State Scale, the sensory domain is assessed directly using the Visual and Auditory Self-Report, the nutritional domain is assessed directly using the Mini-Nutrition Assessment, and the psychological domain is assessed directly using the Geriatric Depression Scale.

[0079] The Mini-Mental State Examination (MMSE) includes tests for orientation, short-term and delayed memory, attention, calculation, naming, retelling, reading, writing and copying, with a score range of 0 to 30.

[0080] The Mini Physical Fitness Scale (SPPB) assesses balance, gait speed, lower limb strength and function, with a score range of 0–12.

[0081] The visual and auditory self-report involves participants selecting the most appropriate answer to report their visual acuity and hearing status: ① Normal (0 points); ② Slightly impaired (1 point); ③ Limited in daily life (2 points); ④ Severely impaired (3 points). The sensory score is the sum of the visual acuity and hearing scores.

[0082] Mini Nutrition Assessment (MNA) includes weight assessment, dietary intake, overall nutritional status, and upper arm and calf circumference measurements, with a score range of 0 to 30.

[0083] The Geriatric Depression Scale (GDS-15) mainly involves 15 questions related to the subject's mood, interests, fatigue, and self-blame, with a score range of 0 to 15.

[0084] The data collection frequencies of different data sources vary greatly. Core scale data may be measured only once every ten or even several weeks in a professional setting, while data on behavior, activity, and physiology generated by wearable devices and mobile terminals are generated almost continuously or in real time, down to the second or minute. This coexistence of fast and slow data makes it impossible to effectively compare, integrate, and model health status across multiple dimensions on the same time frame. This leads to contradictions in the asynchronous description of status. Directly mixing these data may result in temporal misalignment, severely distorting the assessment of the elderly person's overall current internal capabilities, and rendering the assessment results meaningless in terms of timeliness.

[0085] Furthermore, the massive real-time raw data streams (such as heart rate per second) and sparse scale data points are not directly compatible in terms of data structure. Forcibly aligning to low frequencies using downsampling will result in the loss of dynamic patterns in the real-time data; forcibly simulating high-frequency scale data using interpolation will introduce a large amount of noise and false information, and place a huge burden on storage and computation. Therefore, time alignment is necessary to achieve accurate dynamic assessment.

[0086] The generation cycle for the time point T_n closest to the assessment is dynamically set according to the risk level of intrinsic ability (e.g., a short cycle, such as 1-4 days, is used for high risk to closely monitor and intervene in a timely manner; a medium cycle, such as 15-20 days, is used for medium risk to actively follow up and prevent deterioration; a long cycle, such as 1 month, is used for low risk to reduce the burden). For core scale data (which is low-frequency data), the effective assessment result closest to the assessment time point T_n is used. For real-time data (data that needs to be collected at high frequency), a dynamic time window is defined that goes back from T_n. The length of the window is linked to the risk level of intrinsic ability (e.g., the window for high risk is 3 days before T_n, and for low risk it is 30 days before T_n. A mapping table can be set up to read the data directly). Within this window, features are extracted from the real-time data instead of using massive amounts of raw data directly. The dimensions and distribution of the extracted features are different: for each extracted feature, Z-score is used to standardize it so that it follows a standard normal distribution and is normalized to the [0,1] interval.

[0087] The features extracted from real-time data and their indicative meanings are shown in the table below:

[0088] Data categories Extracted features Indicative significance of features Behavioral data Cognitive training: mean accuracy rate, coefficient of variation of reaction time, slope of accuracy rate trend, number of completions. Sensory compensation: average daily operation frequency, average volume level, slope of volume change trend. It objectively reflects the actual loss of cognitive processing ability, stability, and sensory function. Physiological data Sleep: Average deep sleep duration, percentage of days with sleep efficiency <80%, standard deviation of bedtime. Heart: Daily mean HRV, day-night HRV ratio, resting heart rate coefficient of variation (resting heart rate coefficient of variation = (standard deviation of resting heart rate ÷ average of resting heart rate) × 100%). Assess sleep recovery ability, sleep patterns, psychological stress, and autonomic nervous system function. Activity data Activity level: Average daily steps, total standing time. Activity intensity: 75th percentile of walking speed. Activity pattern: Coefficient of variation of walking speed (coefficient of variation of walking speed = (standard deviation of walking speed ÷ average walking speed) × 100%), longest continuous sitting time. Quantify overall activity levels, active performance, gait stability, and sedentary risk. Use the 75th percentile for a more accurate representation of effective activity capacity, avoiding the mean being lowered by low-speed activity. Strength data Current level: The most recent measurement. Trend: The weighted linear regression slope of all historical measurements (with higher weights for recent measurements). By directly quantifying muscle strength and its long-term changes, and using weighted regression, the trend can more accurately reflect recent important changes. Dietary data Compliance: Check-in completion rate. Dietary quality: Food variety score, average daily protein intake frequency. Regularity: Standard deviation of meal check-in time. Assess adherence to nutritional interventions, dietary diversity, and regularity of eating habits. Metabolic data Composition trend: The slope of the linear regression of muscle mass. Composite index: The slope of the trend of the ratio of muscle mass to fat mass. To assess the actual effects of nutritional and exercise interventions on lean body mass, a core composite indicator reflecting improvements in body composition health. Social and behavioral data Activity Level: Average daily social media usage time and average weekly number of communications. Change Alert: Percentage decrease in this week's communication frequency compared to the average of the previous four weeks. Network Size: Number of active contacts. Quantifying social participation and the size of social networks are objective behavioral indicators of mental health.

[0089] Traditional periodic assessments (such as monthly or quarterly) have monitoring blind spots and cannot capture sudden health deterioration events (such as sudden illness or a sharp drop in mobility after a fall) in real time, resulting in delayed warnings and missed optimal intervention periods. This design uses a real-time monitoring channel running in parallel with the main assessment process to detect anomalies in activity and physiological data based on individualized dynamic baselines and introduces multi-indicator cross-validation to achieve "7x24-hour" proactive safety monitoring. It can identify acute risk signs such as a sharp decrease in steps or abnormal heart rate within hours or days and immediately trigger warnings and emergency assessments, transforming post-event response into pre-event or in-event intervention, greatly improving the system's safety assurance capabilities.

[0090] Methods for deploying independent monitoring and response mechanisms for acute events include:

[0091] For each elderly person's activity and physiological data, the system utilizes their historical health data to establish an individualized dynamic baseline for each monitored real-time data point (such as daily steps and resting heart rate). This baseline is not a fixed value, but rather a statistically distributed value that updates over time. The method for obtaining the individualized dynamic baseline is as follows: Data from the most recent N valid days (e.g., N=14) is used as the calculation window. The mean μ and standard deviation σ are calculated for each day's data within this window. For example, the baseline for daily steps is (μ_steps, σ_steps).

[0092] Based on individualized dynamic baselines, dynamic anomaly detection thresholds are set for each indicator, including thresholds for sudden drops or rises in a single day, thresholds for continuous anomalies, and predefined judgment rules.

[0093] Threshold for sudden drop or rise in daily step count: For indicators such as daily step count, a "single-day value below (μ - U × σ)" is considered an abnormal drop. Here, U is a preset sensitivity coefficient (e.g., U = 2.5 or 3.0). This means that when an elderly person's daily step count falls below 2.5 standard deviations of their recent personal average, an anomaly is triggered.

[0094] Continuous abnormality threshold: For indicators such as resting heart rate, two types of abnormalities are defined:

[0095] Absolutely abnormal: Heart rate exceeds the preset absolute safety range (e.g., <40 beats / min or >120 beats / min).

[0096] Relative abnormality: Heart rate values ​​consistently higher than the individual baseline (μ+2σ) for M consecutive hours (e.g., M=24).

[0097] The above decision rules are run in parallel, and cross-validation logic is introduced to improve the specificity of the warning and reduce false alarms.

[0098] Primary trigger: When any metric (such as steps) reaches its dynamic anomaly detection threshold, a primary anomaly signal is generated;

[0099] Cross-validation: Check the status of other indicators within the same time period (e.g., 6 hours before and after) when the primary abnormal signal is generated;

[0100] If at least two indicators trigger primary abnormal signals simultaneously (e.g., a sudden drop in steps in a single day and a continuous abnormal increase in resting heart rate occur at the same time), a high-confidence acute abnormal indicator is immediately generated.

[0101] Single indicator continuous trigger: If a single indicator is abnormal for more than a preset time (e.g., the number of steps is lower than (μ-2σ) for two consecutive days), it will be upgraded to an acute abnormal indicator even without the coordination of other indicators.

[0102] When an acute abnormal indicator is triggered, an acute health risk warning is generated and sent to the preset guardian; the risk level is immediately switched to high risk, and an emergency assessment is arranged according to the first adjustment strategy.

[0103] Standardization using a single population mean (Z-score) ignores the inherent differences in cognitive and motor functions caused by factors such as age, gender, and education level. This leads to systematic biases in assessments of older adults and those with lower levels of education, resulting in insufficient fairness and accuracy. This design addresses this by establishing a stratified reference population database based on age, gender, and education level. Calculations and comparisons are performed within the matched individual's stratum, achieving "homogeneous and comparable" standardization. For example, an 85-year-old's cognitive score is compared to other adults in their age group, rather than to the entire elderly population. This ensures the assessment more accurately reflects their ability level relative to their peers, eliminating confounding factors and making the assessment fairer and more precise.

[0104] The methods for calculating single-domain Z-scores in five domains by combining pre-stored data include:

[0105] We collected core scale data from five domains of a reference healthy older population (referring to a large, representative group of healthy older adults), stratified them according to three key demographic variables: age group, gender, and education level, and calculated the mean and standard deviation of the scores in each domain within each stratum to form a reference population database.

[0106] When assessing a specific elderly person, the system automatically identifies the elderly person's age group, gender, and education level, and uses these as an index to retrieve and match the mean and standard deviation of the corresponding stratum from the reference population database, which are then recorded as pre-stored data.

[0107] For each domain, calculate the single-domain Z-score as follows: (Individual measurement value - mean of the reference population for matching stratification) ÷ standard deviation of the reference population for matching stratification.

[0108] This formula converts an individual's absolute score into a standardized score that measures their relative position against a reference population. The Z-score fluctuates around 0, with a positive score indicating that the individual is above the average level of their age and peers, and a negative score indicating that the individual is below the average level.

[0109] The quality of multi-source data varies greatly (scales may be outdated, and wristbands may not be fully worn). If all data is treated equally during data fusion, low-quality or outdated data will contaminate features, leading to high noise in the model input and unreliable evaluation results. This design assigns a dynamic credibility weight to each data point (based on data completeness and time freshness), and retains this weight information during feature concatenation, constructing a "quality-aware" feature representation. The system can automatically identify and reduce the influence of unreliable data, allowing subsequent model calculations to focus more on high-quality, timely evidence, thereby significantly improving the robustness and anti-interference ability of the overall evaluation system.

[0110] The method for constructing multimodal time-series feature vectors of intrinsic abilities of elderly individuals at multiple assessment time points includes:

[0111] For a single assessment time point, the current assessment time point is used as the anchor point, and the single-domain Z-scores and core scale data of the five domains are directly incorporated; within the dynamic time window, all extracted features are concatenated to form a time-slice multimodal feature vector; for example, the feature vector of the motor domain is [SPPB original scale score, motor domain Z-score, average daily steps, coefficient of variation of gait speed, ...].

[0112] Each data point in the time-slice multimodal feature vector is assigned a dynamic confidence weight (range 0-1). For core scale data, if completed within a dynamic time window specified by T_n, the weight is 1.0; if this period is exceeded, the weight decays over time, but retains the minimum weight (e.g., 0.1). For real-time data, the weight is calculated based on the data completeness rate within the dynamic backtracking window. If the completeness rate is higher than a set completeness threshold (e.g., 80%), the weight is 1.0; otherwise, the weight is reduced proportionally. (For example, if the wristband is worn for ≥80% of the time within the window, the weight is 1.0; otherwise, it is reduced proportionally). This mechanism intelligently assesses and quantifies the reliability of different data sources. In subsequent model calculations, data with higher weights will have a greater influence, thereby improving the robustness of the overall assessment.

[0113] All features in the same domain, after being weighted, are concatenated to form feature sub-vectors for each domain. These sub-vectors are then concatenated to form a complete time-slice multimodal feature vector. Subsequently, the time-slice multimodal feature vectors generated at each evaluation time point are arranged in chronological order according to their corresponding timestamps to obtain an intrinsic capability multimodal time series feature vector. Its structure can be expressed as: [number of time points] × [sum of feature dimensions of the five domains] × [feature value + weight].

[0114] Among them, the data completeness rate refers to the percentage of the actual effective data duration (or number of data points) collected within a specified dynamic backtracking window for a certain real-time data item, relative to the total window duration (or expected number of data points).

[0115] Example of calculation method: If the dynamic backtracking window is 7 days (i.e. 168 hours) before T_n, and the actual duration of effective heart rate data recorded by the bracelet is 120 hours due to charging, not being worn, etc.

[0116] Therefore, the data completeness rate is approximately 71.4% (120 hours ÷ 168 hours).

[0117] Weighting adjustment rules: Set a completeness threshold (e.g., 80%). When the completeness rate is greater than or equal to the completeness threshold, the data quality is considered reliable, and a weight of 1.0 is assigned. When the completeness rate is less than the threshold, the weight decreases linearly, i.e., the decreased weight = data completeness rate ÷ completeness threshold. In the example above, the weight = 71.4% ÷ 80% ≈ 0.89.

[0118] This mechanism quantifies data continuity, ensuring that analysis relies on high-quality, coherent data segments rather than raw data streams filled with interruptions and noise.

[0119] The weight decay of core scale data over time means that for data not generated within the dynamic retrospective window, its timeliness is determined by an effective time threshold, assuming that the threshold is 30 days.

[0120] If the data is fresh: if it was completed within 30 days before T_n, assign it the highest weight = 1.0.

[0121] If the data is outdated: If the completion time is between 30 and 60 days before T_n, the weight decreases linearly from 1.0. Example: weight = max(0.1, 1.0 - (actual age of the data - 30 days) ÷ 30 days). For example, for a scale from 45 days ago, the weight = 1.0 - (45 - 30) ÷ 30 = 0.5. For data older than 60 days, the weight drops to a minimum of 0.1, which is neither completely discarded (it may still have reference value) nor significantly reduced in its influence. This mechanism addresses the inherent delay of low-frequency clinical assessments and automatically reconciles the contradiction between "the latest data is best" and "historical data still has reference value," ensuring that the assessment is based on the current situation while also considering historical trajectories.

[0122] When general machine learning models process such multimodal time-series data, they struggle to simultaneously capture an individual's long-term baseline, inter-domain interactions, and the impact of data quality. This results in insufficient personalized predictions, poor interpretability, and insensitivity to slow decline. This design employs an individualized baseline encoder to learn a "health baseline" from an individual's historical data. Multi-domain deep encoding and multi-scale temporal convolutions are used to model domain characteristics and change patterns at different rates, respectively. Confidence weights are incorporated as biases into the attention mechanism. Temporal consistency regularization loss is used during training, resulting in a "deep understanding" model specifically tailored for assessing the intrinsic abilities of older adults. It not only outputs current scores and future trends but also understands "score changes relative to the individual baseline," "which domains are dominant," and "the confidence level of the prediction," achieving personalized, interpretable, and robust dynamic predictions.

[0123] The intrinsic ability dynamic assessment model includes, in order, three levels: Level 1, Level 2, and Level 3.

[0124] The first level of input consists of multimodal time-series feature vectors, and includes an individualized baseline encoder and a multi-domain deep encoder. The individualized baseline encoder aims to address the issue that Z-scores only reflect the relative position within a population and fail to capture an individual's long-term baseline. It constructs a personal baseline for each elderly person's health status, enabling the assessment to consider both horizontal (compared to the population) and vertical (compared to their own past performance) perspectives. The input consists of the first k time-slice multimodal feature vectors from the multimodal time-series feature vectors (e.g., k=4, representing the initial stable period), and its structure includes domain baseline encoding and temporal aggregation.

[0125] The domain baseline encoding consists of five lightweight encoders with shared weights, each with 32 units and activated by the ReLU function. These encoders process five domain feature sub-vectors at each evaluation time point, outputting the domain baseline encoding. Temporal aggregation is used to input the domain baseline encodings from the G evaluation time points into a unidirectional LSTM in chronological order. The LSTM has 64 hidden units to capture baseline evolution. A 32-unit fully connected layer is then used to generate individualized baseline vectors (32-dimensional). An interpretable "personal health baseline profile" is established, enabling the model to identify risks deviating from individual norms and making it more sensitive to "slow declines."

[0126] The purpose of a multi-domain deep encoder is to perform high-order feature abstraction on the organized domain sub-vectors at each evaluation time point.

[0127] The input is the time slice multimodal feature vector of the current evaluation time point T_n. It consists of five structurally identical but parameter-independent domain coding networks working in parallel. Each network consists of two fully connected layers, set to 64 units and 32 units respectively, activated by the ReLU function. The output is a deep embedding vector (32 dimensions) corresponding to each domain. This step is a deep transformation of the existing information.

[0128] The purpose of the second level is to fuse individual baselines and capture multi-scale dependency patterns of domain embedding vectors over time.

[0129] The output of the first layer is used as input, and the individualized baseline vector is used as a prior condition. It is added to the time slice multimodal feature vector of each evaluation time point T_n (residual connection). Three sets of one-dimensional dilated causal convolutions are used with dilation rates of 1, 2 and 4 and kernel size of 3 to extract patterns at different time spans in parallel. Then, the patterns are input into a two-layer GRU network with 128 and 64 hidden units respectively. The final state of the last GRU layer is taken as the temporal context vector (dimension 64).

[0130] The purpose of the third level is to simulate clinical thinking, dynamically assess the "contribution" of different health domains to the current overall state, and allow the credibility weights of upstream calculations to influence this process.

[0131] The input consists of the time-slice multimodal feature vector and temporal context vector at the current evaluation time point T_n. The feature vectors from the five domains are used as queries, and the temporal context vectors as key-value pairs. Attention is calculated using four heads, each with a dimension of 16. When calculating the attention weights, the average confidence weight of each domain is added as a prior bias to the Softmax calculation, guiding the model to focus more on domains with high data quality. The resulting attention-weighted domain representation is concatenated with the temporal context vector and fused through a gating mechanism, outputting the final comprehensive state representation (64 dimensions). This is then passed through two fully connected layers. One fully connected layer is set to 1 unit and uses a sigmoid function to output a normalized score (0-1), denoted as the individualized comprehensive score of intrinsic ability. The other fully connected layer is set to 2 units and uses a linear function to output two values: a predicted comprehensive score and a probability of decline risk, denoted as the future trend indicator.

[0132] Training methods for dynamic assessment models of intrinsic capabilities include:

[0133] For each elderly person's data, training samples are constructed using a sliding window approach. Each sample contains an input sequence and supervision labels. The input sequence is a time-slice multimodal feature vector of length L, representing consecutive assessment time points. The supervision labels include a current ability label and a future trend label. The current ability label refers to the individualized comprehensive score marked by a domain expert at the last assessment time point of the sequence. The future trend label refers to the supervision signal of the future trend indicator, which is the actual change in the comprehensive score (continuous value) and whether a significant functional decline has occurred (binary classification value) within a dynamic time window after the last assessment time point.

[0134] The total loss function is a weighted sum of the rating regression loss, trend prediction loss, and time-consistency regularization. The rating regression loss is the smooth L1 loss between the output of the individualized comprehensive rating model and the current ability label, which is more robust to outliers. The trend prediction loss is the mean squared error loss between the future trend indicator output by the model and the future trend label. The time-consistency regularization is the L2 norm of the difference in ratings between adjacent time points, which is used to indicate that the model's comprehensive rating output changes smoothly over adjacent time points, avoiding drastic fluctuations.

[0135] The AdamW optimizer is used to iterate training in mini-batch mode. After each batch, the gradient is calculated and the main network parameters are updated through the optimizer. The individualized baseline encoder parameters are kept frozen.

[0136] Training is stopped when the total loss function value on the validation set no longer decreases for several consecutive rounds.

[0137] Deploy the trained model on a cloud platform.

[0138] Risk grading based on a single static threshold (e.g., a score below 0.5 is considered high risk) is mechanical and lagging, failing to reflect the risk difference between "stable at a low level" and "rapid decline from a high level," and also unable to handle the uncertainty of the prediction model itself. This design uses the trend prediction slope (direction and speed) and trend prediction uncertainty (confidence level) as dynamic variables, coupled with the static scoring threshold to form a composite judgment rule, achieving both "proactive" and "prudent" risk assessment. The system can provide early warning based on a clear downward trend before the score falls below the threshold (proactive); simultaneously, when prediction uncertainty is high, it avoids arbitrarily escalating the risk, instead marking it as a state requiring observation (prudent), which highly simulates the decision-making thinking of clinical experts.

[0139] The predefined intrinsic capability risk level is determined by a combination of an individualized comprehensive score and future trend indicators; the future trend indicators include at least the trend prediction slope and the trend prediction uncertainty.

[0140] The determination of a high-risk level requires the following conditions to be met simultaneously: the individualized comprehensive score is lower than the first static threshold, the trend prediction slope is negative, and the trend prediction uncertainty is lower than the preset uncertainty threshold.

[0141] The determination of a medium-risk level meets any of the following conditions:

[0142] Condition 1: The individualized comprehensive score is between the first static threshold and the second static threshold;

[0143] Condition 2: The individualized comprehensive score is higher than the second static threshold, but the trend prediction slope is negative and the trend prediction uncertainty is lower than the uncertainty threshold;

[0144] Condition 3: If the uncertainty of the trend prediction is higher than or equal to the uncertainty threshold, it indicates that the reliability of the prediction is insufficient and it needs to be placed under observation.

[0145] The determination of a low-risk level requires the following conditions to be met simultaneously: the individualized comprehensive score is higher than the second static threshold, and the trend prediction slope is zero or positive.

[0146] The trend prediction slope refers to the average rate of change of the individualized comprehensive score of the elderly's intrinsic abilities within a preset future time window. A negative value (e.g., -0.05 / month) indicates a downward trend in the predicted comprehensive score, which is a core signal of increased risk. A value of zero or close to zero indicates a stable predicted state, while a positive value indicates an improved predicted state. It is used to quantify whether the future will improve or worsen and how quickly it will change, serving as the core basis for dynamic risk assessment.

[0147] Uncertainty refers to the degree of confidence or assurance a model has in predicting the slope of a trend. It reflects the sufficiency of historical data and the clarity of the pattern upon which the prediction is based. A lower value indicates that the model, based on regular and continuous historical data about the elderly person, makes a prediction with high confidence regarding their future trend; a higher value indicates that due to missing data, significant fluctuations, or rare patterns, the model's prediction of the future trend has lower confidence, and the conclusion is more uncertain. It introduces a principle of caution in risk assessment. For example, even if the predicted slope is slightly negative, if the uncertainty is high, the system tends to classify it as medium risk – to be observed – rather than blindly classifying it as low risk. This simulates the prudent thinking of clinical experts, greatly improving the reliability and safety of the system.

[0148] Fixed-frequency assessments and data collection lead to resource waste (over-assessment of stable entities) or risk omissions (under-assessment of deteriorating entities), failing to achieve an optimal balance between assessment efficiency and security. This design links risk levels with parameters such as assessment cycle, dynamic window length, data collection frequency, and completeness threshold, forming differentiated strategies of "enhanced monitoring," "targeted observation," and "energy-saving monitoring," achieving intelligent scheduling of "precise resource allocation." Limited assessment resources and attention are dynamically allocated to high-risk individuals and key areas, while minimizing disruption to low-risk individuals, thus minimizing system operating costs and user burden while ensuring overall security.

[0149] S5 includes:

[0150] The system pre-stores assessment period parameters corresponding to high-risk, medium-risk, and low-risk levels respectively; among them, the high-risk level is associated with the first period parameter, the medium-risk level is associated with the second period parameter, and the low-risk level is associated with the third period parameter, and the first period parameter < the second period parameter < the third period parameter.

[0151] When a high-risk level is determined, the first adjustment strategy is implemented: the period of the next assessment time point is set as the first period parameter, and the dynamic time window length is shortened simultaneously; within the window, the frequency of continuous data collection is increased, and the completeness threshold is reduced, so that it is easier to obtain a higher credibility weight.

[0152] When the risk level is determined to be medium, the second adjustment strategy is implemented: the period of the next assessment time point is set to the second period parameter; based on the attention weight output of the third level of the intrinsic capability dynamic assessment model, the areas with attention weights higher or lower than the historical average are identified, and for the areas, the sampling density of relevant real-time data is increased in the subsequent dynamic time window.

[0153] When the risk level is determined to be low, the third adjustment strategy is implemented: the period for the next assessment time point is set to the third period parameter, while maintaining the existing data collection strategy.

[0154] Example 2, please refer to Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A cloud platform for multimodal dynamic assessment of the intrinsic abilities of the elderly is provided, including:

[0155] Data acquisition and acute monitoring module: used to receive multi-source health data, form a multimodal raw dataset organized according to the assessment time point, and deploy an independent monitoring and handling mechanism for acute events;

[0156] Multi-source data processing module: used to combine pre-stored data to calculate single-domain Z-scores in five domains, and to construct multi-modal time series feature vectors of intrinsic abilities of the elderly under test at multiple assessment time points based on the multimodal raw dataset;

[0157] Evaluation model building and training module: used to deploy dynamic evaluation models of intrinsic capabilities on cloud platforms;

[0158] Dynamic assessment module: used to obtain an individualized comprehensive score of intrinsic ability at the current assessment time point, as well as future trend indicators within a preset time window, based on the intrinsic ability dynamic assessment model;

[0159] Strategy generation module: Used to determine the current intrinsic ability risk level of the elderly based on individualized comprehensive scores and future trend indicators, and adaptively adjust the assessment strategy and data collection mode.

[0160] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0161] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multimodal dynamic assessment method for the intrinsic abilities of the elderly, characterized in that, include: S1. Receive multi-source health data on the cloud platform, perform time alignment, form a multimodal raw dataset organized according to the assessment time point, deploy an independent monitoring and handling mechanism for acute events; and combine with pre-stored data to calculate single-domain Z-scores to represent older adults in five domains: cognition, movement, sensory, nutrition, and psychology. S2. Based on the single-domain Z-scores of five domains and the original multimodal dataset, construct the multimodal time series feature vector of the intrinsic abilities of the elderly under test at multiple assessment time points; S3. Deploy the intrinsic capability dynamic assessment model on the cloud platform, input multimodal time series feature vectors, and then output the individualized comprehensive score of intrinsic capability at the current assessment time point, as well as the future change trend indicators within the future preset time window; The intrinsic ability dynamic assessment model includes, in sequence, a first level, a second level, and a third level: The input of the first level is a multimodal time series feature vector, which includes an individualized baseline encoder and a multi-domain deep encoder. The input of the individualized baseline encoder is the first k time slice multimodal feature vectors in the multimodal time series feature vector. The structure includes domain baseline encoding and temporal aggregation. The domain baseline encoding consists of five lightweight encoders with shared weights, each with 32 units and activated by the ReLU function. Each encoder processes five domain feature sub-vectors at each evaluation time point and outputs the domain baseline encoding. Temporal aggregation is used to input the domain baseline encodings of k evaluation time points into a unidirectional LSTM in chronological order. The LSTM has 64 hidden units and generates individualized baseline vectors through a 32-unit fully connected layer. The input to the multi-domain deep encoder is the time-slice multimodal feature vector at the current evaluation time point T_n. It consists of five domain coding networks with the same structure but independent parameters working in parallel. Each network consists of two fully connected layers, set to 64 units and 32 units respectively, activated by the ReLU function, and the output is a deep embedding vector corresponding to each domain. The second layer takes the output of the first layer as input, the individualized baseline vector as prior condition, and adds it to the time slice multimodal feature vector at each evaluation time point T_n. It uses three sets of one-dimensional dilated causal convolutions with dilation rates of 1, 2 and 4 and a kernel size of 3. Then it is input into a two-layer GRU network with 128 and 64 hidden units respectively. The final state of the last GRU layer is taken as the temporal context vector. The third-level input consists of the time-slice multimodal feature vector and temporal context vector at the current evaluation time point T_n. Five domain feature vectors are used as queries, and the temporal context vector is used as the key. Attention is calculated using four heads, each with a dimension of 16. When calculating the attention weights, the average confidence weight of each domain is added as a prior bias to the Softmax calculation. The resulting attention-weighted domain representation is concatenated with the temporal context vector and fused through a gating mechanism to output the final comprehensive state representation. Then, two fully connected layers are passed. One fully connected layer is set to 1 unit and uses a sigmoid function to output a normalized score, denoted as the individualized comprehensive score of intrinsic ability. The other fully connected layer is set to 2 units and uses a linear function to output two values: a predicted comprehensive score and a probability of decline risk, denoted as the future trend indicator. S4. Predefined intrinsic capability risk level; S5. Based on individualized comprehensive scores and future trend indicators, determine the current intrinsic risk level of the elderly and adaptively adjust the assessment strategy and data collection mode.

2. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 1, characterized in that, The multimodal raw data is used to represent the data that needs to be collected from older adults in five domains, including core scale data and real-time data. The real-time data includes behavioral data, physiological data, activity data, strength data, dietary data, metabolic data, and social and behavioral data. The core scale data includes the raw scale scores for each domain. The cognitive domain is assessed directly using the Mini-Mental State Scale, the motor domain is assessed directly using the Mini-Mental State Scale, the sensory domain is assessed directly using the Visual and Auditory Self-Report, the nutritional domain is assessed directly using the Mini-Nutrition Assessment, and the psychological domain is assessed directly using the Geriatric Depression Scale. The generation cycle of the nearest assessment time point T_n is dynamically set according to the intrinsic ability risk level. For core scale data, the effective assessment result closest to the assessment time point T_n is used. For real-time data, a dynamic time window is defined that goes back from T_n. The length of the window is linked to the intrinsic ability risk level. Within this window, features are extracted from the real-time data. For each extracted feature, Z-score is used for standardization and normalization to the [0,1] interval.

3. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 2, characterized in that, The methods for deploying an independent monitoring and response mechanism for acute events include: For each elderly person, an individualized dynamic baseline is established based on their activity and physiological data; this baseline is a statistical distribution value that is updated over time. Based on an individualized dynamic baseline, dynamic anomaly thresholds are set for each indicator, including thresholds for sudden drops or rises in a single day and thresholds for continuous anomalies. Predefined judgment rules are established, and cross-validation logic is introduced. When any indicator reaches its dynamic anomaly threshold, a primary anomaly signal is generated. During the same time period when the primary anomaly signal is generated, the status of other indicators is checked. If at least two indicators trigger primary anomaly signals simultaneously, a high-confidence acute anomaly indicator is immediately generated. If a single indicator remains abnormal for more than a preset duration, it is upgraded to an acute anomaly indicator. When an acute anomaly indicator is triggered, an acute health risk warning is generated and sent to the preset monitoring party. The risk level is immediately switched to high risk, and an emergency assessment is arranged according to the first adjustment strategy.

4. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 3, characterized in that, The methods for calculating single-domain Z-scores in five domains by combining pre-stored data include: Collect data from five core scales in the reference healthy older adults; The core scale data was stratified according to three key demographic variables: age group, gender, and education level. Within each stratum, the mean and standard deviation of the scores in the five domains were calculated to form a reference population database. When assessing a specific elderly person, the system automatically identifies the elderly person's age group, gender, and education level, and uses these as an index to retrieve and match the mean and standard deviation of the corresponding stratum from the reference population database, which are then recorded as pre-stored data. For each domain, calculate the single-domain Z-score as follows: (Individual measurement value - mean of the reference population for matching stratification) ÷ standard deviation of the reference population for matching stratification.

5. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 4, characterized in that, The method for constructing a multimodal time-series feature vector of the intrinsic abilities of the elderly under test at multiple assessment time points includes: For a single assessment time point, using the current assessment time point as the anchor point, the single-domain Z-scores and core scale data of the five domains are directly incorporated; Within the dynamic time window, each feature data is assigned a dynamic credibility weight. For core scale data, if it is completed within the dynamic time window specified by T_n, the weight is 1.

0. If it exceeds this dynamic time window, the weight decays over time, but the minimum weight is retained. For real-time data, the weight is calculated based on the data completeness rate within the dynamic time window. If the completeness rate is higher than the set completeness threshold, the weight is 1.

0. If it is lower than the threshold, the weight is reduced proportionally. All features in the same domain after being weighted are concatenated to form feature sub-vectors for each domain. Then, the feature sub-vectors for each domain are concatenated to form a time-slice multimodal feature vector corresponding to the evaluation time point. Subsequently, the time-slice multimodal feature vectors generated at each evaluation time point are arranged in chronological order according to their corresponding timestamps to obtain an intrinsic capability multimodal time series feature vector.

6. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 5, characterized in that, The training method for the intrinsic ability dynamic assessment model includes: For each elderly person's data, training samples are constructed using a sliding window approach. Each sample contains an input sequence and supervision labels. The input sequence is a time-slice multimodal feature vector of length L with consecutive assessment time points. The supervision labels include current ability labels and future trend labels. The current ability label refers to the individualized comprehensive score marked by a domain expert at the last assessment time point of the sequence. The future trend label refers to the supervision signal of the actual comprehensive score change and whether a significant functional decline has occurred within a dynamic time window after the last assessment time point. The total loss function is a weighted sum of rating regression loss, trend prediction loss, and temporal consistency regularization. Rating regression loss refers to the smoothing L1 loss between the output of the individualized comprehensive rating model and the current ability label; trend prediction loss refers to the mean squared error loss between the future trend indicator output by the model and the future trend label; and temporal consistency regularization refers to the L2 norm of the rating difference between adjacent time points. The AdamW optimizer is used to iterate training in mini-batch mode. After each batch, the gradient is calculated and the main network parameters are updated through the optimizer. The individualized baseline encoder parameters are kept frozen. Training is stopped when the total loss function value on the validation set no longer decreases for several consecutive rounds. Deploy the trained model on a cloud platform.

7. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 6, characterized in that, The predefined intrinsic capability risk level is determined by a combination of an individualized comprehensive score and a future trend indicator; the future trend indicator includes at least the trend prediction slope and the trend prediction uncertainty. The determination of a high-risk level requires the following conditions to be met simultaneously: the individualized comprehensive score is lower than the first static threshold, the trend prediction slope is negative, and the trend prediction uncertainty is lower than the preset uncertainty threshold. The determination of a medium-risk level meets any of the following conditions: Condition 1: The individualized comprehensive score is between the first static threshold and the second static threshold; Condition 2: The individualized comprehensive score is higher than the second static threshold, but the trend prediction slope is negative and the trend prediction uncertainty is lower than the uncertainty threshold; Condition 3: The uncertainty of trend prediction is higher than or equal to the uncertainty threshold; The determination of a low-risk level requires the following conditions to be met simultaneously: the individualized comprehensive score is higher than the second static threshold, and the trend prediction slope is zero or positive. Among them, the trend prediction slope refers to the average rate of change of the individualized comprehensive score of the intrinsic ability of the elderly within a preset time window in the future, and the uncertainty of change refers to the degree of grasp or confidence level of the model on the trend prediction slope.

8. The multimodal dynamic assessment method for intrinsic abilities of the elderly according to claim 7, characterized in that, S5 includes: The system pre-stores assessment period parameters corresponding to high-risk, medium-risk, and low-risk levels respectively; among them, the high-risk level is associated with the first period parameter, the medium-risk level is associated with the second period parameter, and the low-risk level is associated with the third period parameter, and the first period parameter < the second period parameter < the third period parameter. When a high-risk level is determined, the first adjustment strategy is implemented: the period of the next assessment time point is set to the first period parameter, and the dynamic time window length is shortened simultaneously; within the window, the frequency of continuous data collection is increased, and the completeness threshold is reduced. When the risk level is determined to be medium, the second adjustment strategy is implemented: the period of the next assessment time point is set to the second period parameter; based on the attention weight output of the third level of the intrinsic capability dynamic assessment model, the areas with attention weights higher than the historical average are identified, and for these areas, the sampling density of relevant real-time data is increased in the subsequent dynamic time window. When the risk level is determined to be low, the third adjustment strategy is implemented: the period for the next assessment time point is set to the third period parameter, while maintaining the existing data collection strategy.

9. A cloud platform for multimodal dynamic assessment of intrinsic abilities of the elderly, used to implement the multimodal dynamic assessment method for intrinsic abilities of the elderly as described in any one of claims 1 to 8, characterized in that, include: Data acquisition and acute monitoring module: used to receive multi-source health data, form a multimodal raw dataset organized according to the assessment time point, and deploy an independent monitoring and handling mechanism for acute events; Multi-source data processing module: used to combine pre-stored data to calculate single-domain Z-scores in five domains, and to construct multi-modal time series feature vectors of intrinsic abilities of the elderly under test at multiple assessment time points based on the multimodal raw dataset; Evaluation model building and training module: used to deploy dynamic evaluation models of intrinsic capabilities on cloud platforms; Dynamic assessment module: used to obtain an individualized comprehensive score of intrinsic ability at the current assessment time point, as well as future trend indicators within a preset time window, based on the intrinsic ability dynamic assessment model; Strategy generation module: Used to determine the current intrinsic ability risk level of the elderly based on individualized comprehensive scores and future trend indicators, and adaptively adjust the assessment strategy and data collection mode.

Images