A question-and-answer-based elderly health management method and device

Abstract

This invention discloses a question-and-answer-based elderly health management method and apparatus, relating to the field of elderly health management technology. The method includes constructing a cognitive characteristic database of the elderly population, building a dynamic question-and-answer strategy library, collecting physiological, cognitive, and sensory parameters during peak daily activity periods of the elderly, calculating health coefficients corresponding to each dimension, outputting a comprehensive health score using neural attention mechanism technology, classifying health levels, and providing feedback reports to update the strategy library. The apparatus includes a dynamic question-and-answer module, a rhythm perception scheduling module, a multimodal data acquisition module, a health coefficient calculation module, an adaptive fusion calculation module, a health level determination module, and a closed-loop feedback optimization module. This invention improves the accuracy of health assessment and effectively solves the technical problems of traditional elderly health management lacking specificity, failing to fully utilize multi-source data, and being unable to dynamically adjust management strategies.

Description

Technical Field

[0001] This invention belongs to the field of elderly health management technology, specifically relating to a question-and-answer type elderly health management method and device. Background Technology

[0002] With the acceleration of global aging, health management for the elderly has become a focus of social attention.

[0003] While existing elderly health management methods have achieved some degree of health management, they suffer from numerous problems. Firstly, data collection typically employs timed or random sampling, failing to consider the diurnal rhythms of cognitive and physiological states in the elderly. This results in insufficient data representativeness and an inability to accurately reflect the true health status of seniors. For instance, the elderly's responses to cognitive tests and the stability of physiological indicators vary across different time periods, and traditional collection methods easily overlook these crucial pieces of information. Secondly, health assessments often focus on single or a few physiological parameters, such as blood pressure and blood sugar, neglecting the significant impact of cognitive and sensory functions on the overall health of the elderly. Furthermore, the weights of each dimension in the assessment process are often fixed, failing to adapt to the dynamic changes in the importance of each health dimension at different stages for different individuals. Moreover, existing methods lack effective utilization of the temporal characteristics of historical health data, making it difficult to identify potential trends in health status. Finally, the entire health management process is unidirectional, lacking a mechanism for dynamically optimizing management strategies based on feedback information, thus failing to meet the complex and ever-changing health needs of the elderly.

[0004] Therefore, a question-and-answer-based elderly health management method and device have emerged. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a question-and-answer type elderly health management method and apparatus to solve the following technical problem:

[0006] On the one hand, data collection typically employs timed or random sampling, failing to consider the diurnal rhythm variations in cognitive and physiological states of the elderly, resulting in insufficient data representativeness and an inability to accurately reflect their true health status. On the other hand, health assessments often focus on single or a few physiological parameters, such as blood pressure and blood sugar, neglecting the significant impact of cognitive and sensory functions on the overall health of the elderly. Furthermore, the weights of each dimension in the assessment process are often fixed, failing to adapt to the dynamic changes in the importance of each health dimension at different stages for different individuals. Moreover, existing methods lack effective utilization of the temporal characteristics of historical health data, making it difficult to identify potential trends in health status. In addition, the entire health management process is unidirectional, lacking a mechanism for dynamically optimizing management strategies based on feedback information, thus failing to meet the complex and ever-changing health needs of the elderly.

[0007] To address the above problems, a first aspect of the present invention provides a question-and-answer-based elderly health management method, comprising the following steps:

[0008] S1: Construct a cognitive characteristic database of the elderly population and a dynamic question-answering strategy library, which includes an initial question sequence and question adjustment rules;

[0009] The steps for constructing the dynamic question-answering strategy library are as follows: extract core scoring criteria and typical questions based on the MMSE scale, convert them into quantifiable data labels through keyword matching, establish a triplet model, and form a dynamic question-answering strategy library by setting a dynamic adjustment mechanism.

[0010] S2: Based on the rhythm analysis results in the cognitive characteristics database of the elderly population, health data collection for the elderly should be initiated during the peak periods of their daily activity levels.

[0011] S3: Simultaneously collect physiological, cognitive, and sensory parameters of the elderly;

[0012] S4: For the three dimensions of physiology, cognition and senses, the corresponding health coefficients for each dimension are generated through interval mapping, standardized testing and normalization, including the physiological health coefficient, cognitive health coefficient and sensory health coefficient of the elderly.

[0013] S5: Adaptively allocate cross-dimensional weights using the neural attention mechanism, input the historically accumulated health data sequence into the gated recurrent unit, extract temporal dependent features and fuse them with the current interaction item, perform confidence verification through a dual-path machine learning model, and finally output a normalized comprehensive health score.

[0014] S6: Analyze the complete health information obtained, establish an elderly health indicator system, perform Z-score standardization, use the model to output the comprehensive health score of the elderly, and classify the health level;

[0015] S7: Provide health assessment reports to the elderly and their guardians, and update the dynamic question-and-answer strategy database based on the feedback.

[0016] Preferably, step S1 includes the following steps:

[0017] The elderly population cognitive characteristics database includes the elderly's memory level, language comprehension ability, reaction speed, and seasonal illness attack patterns.

[0018] The question adjustment rules are formulated based on the semantic similarity of the answer information, the answer duration, the health language and number comprehension deviation index, and the diurnal rhythm characteristics.

[0019] The circadian rhythm characteristics were determined using time series analysis to calculate the daily peak periods, trough periods, and other time periods for the elderly.

[0020] Preferably, step S2 includes the following steps:

[0021] Based on the circadian rhythm characteristics in the cognitive characteristic database of the elderly population, the data collection work was started when the daily activity level of the elderly was at its peak.

[0022] During the data collection process, elderly health data from various dimensions are recorded in real time, including physiological parameters, cognitive parameters, and sensory parameters.

[0023] The physiological parameters include blood pressure, blood sugar, heart rate, and respiratory rate. The physiological health coefficient of the elderly is calculated based on the physiological parameters.

[0024] The cognitive parameters include memory level, language comprehension ability, reaction speed, and health language and math comprehension deviation coefficient. The cognitive health coefficient of the elderly is calculated based on the cognitive parameters.

[0025] The sensory parameters include visual parameters, auditory parameters, and proprioceptive positioning accuracy. The sensory health coefficient of the elderly is calculated based on the sensory parameters.

[0026] Preferably, the physiological health coefficient of the elderly includes the following steps:

[0027] For four physiological parameters, age-stratified health reference intervals were established for the elderly. The measured values ​​were converted into corresponding physiological health indices by interval mapping. Blood pressure was calculated by standard deviation weighting, while the other parameters were directly mapped.

[0028] According to the preset weight allocation design, the four physiological health indices are weighted and summed.

[0029] By introducing a metabolic age compensation factor to correct the weighted results, the physiological health coefficient of the elderly is obtained.

[0030] The health reference interval for age stratification is determined based on the percentile of the elderly population in the same age group; the standard deviation weighting uses the coefficient of variation as the weighting coefficient; the metabolic age compensation factor is calculated by the ratio of actual physiological age measured by biomarkers to calendar age.

[0031] Preferably, the cognitive health coefficient of the elderly includes the following steps:

[0032] The standardized value of delayed recall test score was used as the memory level, the accuracy of semantic matching task was used as the language comprehension ability, the reaction speed was converted into the reaction efficiency value, which was obtained by dividing the actual reaction time by the age-matched benchmark reaction time, and the health language and math comprehension deviation index was determined by pre-testing to determine the cognitive error of professional vocabulary among the elderly.

[0033] The corresponding weights are determined by combining the Delphi method with feature importance analysis using machine learning, and then weighted summation is performed to obtain the cognitive health coefficient of the elderly. Specifically:

[0034]

[0035] in, The cognitive health index of the elderly For memory level, For language comprehension ability, For reaction rate, For the health-related language and math comprehension deviation index, , , and These are the corresponding weights.

[0036] Preferably, the sensory health coefficient of the elderly includes the following steps:

[0037] Visual parameters, auditory parameters, and proprioceptive positioning accuracy parameters are obtained. The proprioceptive positioning accuracy is the ratio of the actual deviation angle to the theoretical value when an elderly person completes a joint angle reproduction task with their eyes closed.

[0038] The visual and auditory parameters were normalized separately.

[0039] The corresponding weights were determined using the analytic hierarchy process (AHP), and a weighted fusion model was constructed to obtain the sensory health coefficient of the elderly, specifically:

[0040]

[0041] in, The sensory health index of the elderly. and These are the normalized visual and auditory parameters, respectively. This refers to the proprioceptive positioning accuracy parameter. , and These are the corresponding weights.

[0042] Preferably, step S5 includes the following steps:

[0043] The physiological health coefficient, cognitive health coefficient, and sensory health coefficient of the elderly are used as input vectors. ,in, The physiological health index of the elderly. The cognitive health index of the elderly Sensory health coefficient for the elderly;

[0044] The weights of each dimension are dynamically allocated using a neural attention mechanism.

[0045] The time-accumulated health data sequence is subjected to memory enhancement processing through a gated recurrent unit, and the output hidden state and the Hadamard product of the current input are used as feature interaction terms.

[0046] Random forest and XGBoost models are run in parallel, and the Bhattacharyya distance between the two models' predicted probability distributions is taken as the final confidence score.

[0047] The weighted composite index is normalized to the [0, 100] interval using the Sigmoid function to obtain the final comprehensive health score.

[0048] Preferably, the dynamic allocation of weights for each dimension using a neural attention mechanism includes the following steps:

[0049] The weights of each dimension satisfy... ,and ,in, The historical contribution attenuation factor for the corresponding dimension;

[0050] The dynamic update process of the historical contribution decay factor is as follows:

[0051] For each health dimension Set the initial decay factor and define the forgetting factor;

[0052] After each backpropagation, calculate the partial derivative of the loss function with respect to the input features;

[0053] The decay factors for each dimension are refreshed iteratively based on exponential smoothing, specifically as follows:

[0054]

[0055] in, The updated decay factor, The historical contribution decay factor, i.e., the record cutoff time. At that time, the first The historical cumulative importance score of each feature Forgetting Gating Coefficient For the current loss function on the th Partial derivatives of each feature;

[0056] After all updates After normalization using the Softmax function, the dynamic weights for the next training cycle are obtained;

[0057] If a certain dimension in three consecutive iterations ,in, If the preset threshold is used, the learning rate for that dimension will be forcibly reset to the default value.

[0058] Preferably, in step S6, classifying health levels includes the following steps:

[0059] The health level is divided into five levels based on the overall health score, as follows:

[0060] A health score of ≥85 is considered excellent; 70≤health score≤84 is considered good; 55≤health score≤69 is considered average; 40≤health score≤54 is considered poor; and a health score <40 is considered very poor.

[0061] The risk threshold was determined based on longitudinal tracking data of people in the same age group. The survival curve was fitted using the Weibull distribution, and the 10th percentile was used as the high-risk threshold.

[0062] Each level corresponds to a different management agreement, including monitoring frequency, medication reminders, and emergency contact linkage mechanisms.

[0063] A second aspect of the present invention provides a question-and-answer type elderly health management method device, comprising:

[0064] Dynamic question answering module: Constructs a cognitive characteristic database of the elderly population and a dynamic question answering strategy library, which includes an initial question sequence and question adjustment rules;

[0065] Rhythm-aware scheduling module: Based on the rhythm analysis results in the cognitive characteristic database of the elderly population, prioritize the collection of health data of the elderly during the peak period of their daily activity.

[0066] Multimodal data acquisition module: Simultaneously collects physiological, cognitive, and sensory parameters of the elderly;

[0067] Health coefficient calculation module: For the three dimensions of physiology, cognition and senses, it generates the corresponding health coefficients for each dimension through interval mapping, standardized testing and normalization, including the physiological health coefficient, cognitive health coefficient and sensory health coefficient of the elderly.

[0068] Adaptive fusion computing module: It uses the neural attention mechanism to adaptively allocate cross-dimensional weights, inputs the historical accumulated health data sequence into the gated recurrent unit, extracts the time-dependent features and fuses them with the current interaction item, performs confidence verification through a dual-path machine learning model, and finally outputs a normalized comprehensive health score.

[0069] Health level assessment module: Analyzes the complete health information obtained, establishes an elderly health indicator system, performs Z-score standardization, uses the model to output the comprehensive health score of the elderly, and classifies the health level;

[0070] Closed-loop feedback optimization module: The health assessment report is fed back to the elderly and their relevant guardians, and the dynamic question-and-answer strategy library is updated based on the feedback information.

[0071] The beneficial effects of this invention are:

[0072] This invention uses the circadian rhythm characteristics analysis results from the cognitive characteristic database of the elderly population to collect health data during the peak daily activity periods of the elderly population. This can obtain more representative and accurate health data, providing a solid data foundation for subsequent health level assessment and improving the reliability of health assessment for the elderly.

[0073] This invention constructs a health assessment system encompassing three dimensions: physiological, cognitive, and sensory, comprehensively reflecting the health status of the elderly. By integrating and weighting multiple parameters within each dimension, a corresponding health coefficient is obtained. Compared to traditional single-dimensional assessments, this system can more accurately pinpoint health problems in the elderly.

[0074] This invention adaptively allocates cross-dimensional weights through a neural attention mechanism, combines a gated recurrent unit to fuse historical and current health data, and outputs a comprehensive health score. This makes the health assessment results more consistent with the actual health changes of the elderly, improving the timeliness and accuracy of the assessment. At the same time, it divides the health into five levels based on the comprehensive health score, determines the level threshold based on longitudinal tracking data of people of the same age group, and formulates differentiated management protocols for each level. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0076] Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0077] 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.

[0078] Please see Figure 1 As shown, this invention is a question-and-answer-based method for elderly health management, comprising the following steps:

[0079] S1: Construct a cognitive characteristic database of the elderly population and a dynamic question-answering strategy library, which includes an initial question sequence and question adjustment rules;

[0080] The steps for constructing the dynamic question-answering strategy library are as follows: extract core scoring criteria and typical questions based on the MMSE scale, convert them into quantifiable data labels through keyword matching, establish a triplet model, and form a dynamic question-answering strategy library by setting a dynamic adjustment mechanism.

[0081] S2: Based on the rhythm analysis results in the cognitive characteristics database of the elderly population, health data collection for the elderly should be initiated during the peak periods of their daily activity levels.

[0082] S3: Simultaneously collect physiological, cognitive, and sensory parameters of the elderly;

[0083] S4: For the three dimensions of physiology, cognition and senses, the corresponding health coefficients for each dimension are generated through interval mapping, standardized testing and normalization, including the physiological health coefficient, cognitive health coefficient and sensory health coefficient of the elderly.

[0084] S5: Adaptively allocate cross-dimensional weights using the neural attention mechanism, input the historically accumulated health data sequence into the gated recurrent unit, extract temporal dependent features and fuse them with the current interaction item, perform confidence verification through a dual-path machine learning model, and finally output a normalized comprehensive health score.

[0085] S6: Analyze the complete health information obtained, establish an elderly health indicator system, perform Z-score standardization, use the model to output the comprehensive health score of the elderly, and classify the health level;

[0086] S7: Provide health assessment reports to the elderly and their guardians, and update the dynamic question-and-answer strategy database based on the feedback.

[0087] Specifically, data on memory and language abilities are extracted from a cognitive characteristic database of the elderly population. An initial question sequence is constructed using the MMSE scale. Scoring criteria and typical questions for memory and language abilities are extracted from the MMSE scale and converted into quantifiable data labels (keyword matching). A tripartite model of question type-cognitive domain-difficulty level is established. Seven basic question types are extracted and refined from the MMSE (time orientation, place orientation, immediate memory, delayed recall, naming ability, retelling ability, and reading comprehension). The correspondence between these types and brain region functions is established using the NINCDS-ADRDA diagnostic criteria. The difficulty coefficient is calculated by averaging the quotient of the number of question words to the average number of question words in the overall population and the quotient of the elderly person's reaction time to the average reaction time of the same type of elderly person. The calculation formula is as follows: ,in, Difficulty level For the number of words in the question, The average number of words in the overall question vocabulary. For the reaction time of the elderly, The system uses behavioral data such as the accuracy of user responses and Bayesian inference algorithms to dynamically adjust the current cognitive level profile of elderly individuals (based on average reaction time for similar age groups) to form a dynamic question-and-answer strategy library. By dynamically adjusting the difficulty and order of questions—increasing the difficulty of the next question after three consecutive correct answers and decreasing the difficulty of subsequent questions after two consecutive incorrect answers—the system can sensitively detect early signs of decline that may be missed by traditional static tests. Based on the circadian rhythm analysis results in the database, data collection is initiated during the peak activity period of the elderly, simultaneously collecting physiological, cognitive, and sensory parameters. Physiological parameters are used to generate coefficients using interval mapping, cognitive parameters are standardized, and sensory parameters are normalized to obtain health coefficients for each dimension. The neural attention mechanism is used to assign cross-dimensional weights, and historical data is processed using a gated loop unit. After dual-path model verification, a comprehensive health score is output and normalized. Health information is analyzed to establish an indicator system. After Z-score standardization, five health levels are divided based on the comprehensive score, generating an assessment report containing health level, risk warnings, and intervention suggestions. This report is pushed to the elderly and their guardians via an app, collecting feedback, including understanding of the report and suggestions for question expression, and updating the dynamic question-and-answer strategy library to form a closed-loop optimization.

[0088] In one embodiment of the present invention, step S1 includes the following steps:

[0089] The elderly population cognitive characteristics database includes the elderly's memory level, language comprehension ability, reaction speed, and seasonal illness attack patterns.

[0090] The question adjustment rules are formulated based on the semantic similarity of the answer information, the answer duration, the health language and number comprehension deviation index, and the diurnal rhythm characteristics.

[0091] The circadian rhythm characteristics were determined using time series analysis to calculate the daily peak periods, trough periods, and other time periods for the elderly.

[0092] Specifically, a sample of elderly individuals aged 60-89 without major underlying diseases was selected. Stratified sampling was used to cover different age groups (60-64, 65-69, 70-74, 75-79, 80-84, and 85-89), education levels (primary school and below, junior high school and senior high school and above), and health status (healthy, stable chronic disease period, acute chronic disease period). Data was collected through one-on-one home visits, using standardized scales and smart devices for a 7-day period. A seasonal illness flare-up database was designed using a seasonal health log to help the elderly... The elderly individuals or their guardians recorded common ailments in different seasons over the past three years, including frequency of onset, duration, and triggering factors, and established a correlation database. The extraction of circadian rhythm characteristics was achieved through activity data collection. The elderly individuals in the sample wore smart bracelets and their daily activities were recorded for 7 consecutive days from 0 to 24 hours. The sampling frequency was 5 minutes per time. Time series analysis was used to calculate the daily peak activity periods (continuous periods with steps ≥ 60% of the average daily steps). The common peak periods for 80% of the elderly individuals in the sample were "7-9 am" and "15-17 pm", while the underestimated period was "22-6 am the next day".

[0093] In one embodiment of the present invention, step S2 includes the following steps:

[0094] Based on the circadian rhythm characteristics in the cognitive characteristic database of the elderly population, the data collection work was started when the daily activity level of the elderly was at its peak.

[0095] During the data collection process, elderly health data from various dimensions are recorded in real time, including physiological parameters, cognitive parameters, and sensory parameters.

[0096] The physiological parameters include blood pressure, blood sugar, heart rate, and respiratory rate. The physiological health coefficient of the elderly is calculated based on the physiological parameters.

[0097] The cognitive parameters include memory level, language comprehension ability, reaction speed, and health language and math comprehension deviation coefficient. The cognitive health coefficient of the elderly is calculated based on the cognitive parameters.

[0098] The sensory parameters include visual parameters, auditory parameters, and proprioceptive positioning accuracy. The sensory health coefficient of the elderly is calculated based on the sensory parameters.

[0099] In one embodiment of the present invention, the physiological health coefficient of the elderly includes the following steps:

[0100] For four physiological parameters, age-stratified health reference intervals were established for the elderly. The measured values ​​were converted into corresponding physiological health indices by interval mapping. Blood pressure was calculated by standard deviation weighting, while the other parameters were directly mapped.

[0101] According to the preset weight allocation design, the four physiological health indices are weighted and summed.

[0102] By introducing a metabolic age compensation factor to correct the weighted results, the physiological health coefficient of the elderly is obtained.

[0103] The health reference interval for age stratification is determined based on the percentile of the elderly population in the same age group; the standard deviation weighting uses the coefficient of variation as the weighting coefficient; the metabolic age compensation factor is calculated by the ratio of actual physiological age measured by biomarkers to calendar age.

[0104] Specifically, for each stratum of samples, the 5th percentile (P5) and 95th percentile (P95) of the four physiological parameters are calculated, and [P5, P95] is used as the health reference interval for that age group. For example, the reference interval for systolic blood pressure for men aged 60-64 is [120, 150] mmHg, determined based on the distribution of 90% of the health data in that stratum of samples. The coefficients of variation of systolic and diastolic blood pressure are calculated as the basis for weight allocation, specifically: , ,in, and These are the coefficients of variation for systolic and diastolic blood pressure, respectively. and These are the standard deviations of systolic and diastolic blood pressure for the same age group. and These are the mean values; the systolic blood pressure weighting coefficient is obtained by dividing the systolic blood pressure coefficient of variation by the sum of the systolic and diastolic blood pressure coefficients of variation, and the diastolic blood pressure weighting coefficient is obtained using the same method; the values ​​are then converted into individual indices through interval mapping, mapping systolic and diastolic blood pressure to [0, 100] respectively, and then weighted and fused, specifically as follows:

[0105]

[0106] in, This is the mapped blood pressure index. and These are the measured systolic and diastolic blood pressures, respectively. and These are the systolic and diastolic blood pressures corresponding to the 5th percentile, respectively. and These are the systolic and diastolic blood pressures corresponding to the 95th percentile, respectively. and The weighting coefficients for systolic and diastolic blood pressure are 0.5 each; the remaining indices, such as blood glucose, heart rate, and respiratory rate, are directly mapped. The weights of the four physiological indices are determined using the Delphi method: blood pressure (0.3), blood glucose (0.25), heart rate (0.25), and respiratory rate (0.2). These weighted indices are then summed to obtain the initial physiological health coefficient for the elderly. The metabolic age compensation factor is then calculated as follows: ,in, To determine the actual physiological age using biomarkers, The initial physiological health coefficient of the elderly is adjusted based on calendar age to obtain the adjusted physiological health coefficient, specifically: ,in, This represents the initial physiological health coefficient of the elderly. It is a metabolic age compensating factor; among which, when hour, ;like ,but A value less than 1 indicates an increase in physiological health; otherwise, it indicates a decrease. This reflects the degree of matching between physiological age and actual physical condition.

[0107] In one embodiment of the present invention, the cognitive health coefficient of the elderly includes the following steps:

[0108] The standardized value of delayed recall test score was used as the memory level, the accuracy of semantic matching task was used as the language comprehension ability, the reaction speed was converted into the reaction efficiency value, which was obtained by dividing the actual reaction time by the age-matched benchmark reaction time, and the health language and math comprehension deviation index was determined by pre-testing to determine the cognitive error of professional vocabulary among the elderly.

[0109] The corresponding weights are determined by combining the Delphi method with feature importance analysis using machine learning, and then weighted summation is performed to obtain the cognitive health coefficient of the elderly. Specifically:

[0110]

[0111] in, The cognitive health index of the elderly For memory level, For language comprehension ability, For reaction rate, For the health-related language and math comprehension deviation index, , , and These are the corresponding weights.

[0112] Specifically, a delayed recall test was used, presenting the elderly with 10 everyday words, giving them one minute to memorize them, followed by distractions with other unrelated tasks. After 10 minutes, they were asked to recall the words, and the number of correctly recalled words was recorded. A semantic matching task was also used, providing 20 word pairs, of which 10... Ten groups were synonymous, and 10 groups were non-synonymous. The elderly participants were asked to determine whether each group was synonymous and the number of correct judgments was recorded. A number Stroop test was used, presenting 30 colored numbers. The participants were asked to quickly say the color of the number (not the number itself), and the total reaction time was recorded. The average reaction time for the same age group was used, and the result was obtained by dividing the average average reaction time by the measured reaction time. The Health Language and Number Comprehension Deviation Index was determined through a pre-test design containing 30 common health-related terms, each with three options (correct explanation, ambiguous explanation, and incorrect explanation). The elderly participants completed the questionnaire independently, and the selection results for each term were recorded. The number of incorrectly understood terms was counted, and the Health Language and Number Comprehension Deviation Index was calculated. A higher value indicates a more severe cognitive deviation. Five to seven geriatric medicine experts were invited to score the importance of four parameters. The average score was used as the initial weight, and machine learning feature importance analysis was combined: historical cognitive assessment data was input into a random forest model, and the feature importance scores for each parameter were output. This score was then weighted and fused with the expert scores (weight ratio: expert 60%, model 40%) to determine the final index. , , and The corresponding weight values ​​are 0.3, 0.25, 0.25 and 0.2, respectively.

[0113] In one embodiment of the present invention, the sensory health coefficient of the elderly includes the following steps:

[0114] Visual parameters, auditory parameters, and proprioceptive positioning accuracy parameters are obtained. The proprioceptive positioning accuracy is the ratio of the actual deviation angle to the theoretical value when an elderly person completes a joint angle reproduction task with their eyes closed.

[0115] The visual and auditory parameters were normalized separately.

[0116] The corresponding weights were determined using the analytic hierarchy process (AHP), and a weighted fusion model was constructed to obtain the sensory health coefficient of the elderly, specifically:

[0117]

[0118] in, The sensory health index of the elderly. and These are the normalized visual and auditory parameters, respectively. This refers to the proprioceptive positioning accuracy parameter. , and These are the corresponding weights.

[0119] Specifically, visual parameters are obtained by combining a visual acuity chart with a contrast test. The process is as follows: Visual acuity test: Using an international standard visual acuity chart, the visual acuity value corresponding to the smallest optotype that the elderly person can clearly identify (e.g., 0.8 or 1.0) is recorded as V1. Ten sets of images with different gray levels (contrast from 10% to 100%) are displayed, and the lowest contrast value that can be accurately identified (e.g., 30%) is recorded as V2. The visual parameters are obtained by weighted summation. V1 represents the visual acuity value corresponding to the smallest optotype that the elderly person can clearly identify, and V2 represents the lowest contrast value that the elderly person can accurately identify. Auditory parameters were obtained through pure-tone audiometry, measuring hearing thresholds at three frequencies: 500Hz, 1000Hz, and 2000Hz. The thresholds were quantified according to the following criteria: ≤25dB average hearing threshold = 1 point; 26-40dB = 0.6 points; 41-60dB = 0.3 points; >60dB = 0.1 points. Proprioceptive positioning accuracy parameters were obtained through a joint angle reproduction task, testing the knee and elbow joints. For each joint: the tester passively moves the joint to the target angle, allowing the elderly person to perceive it for 10 seconds before returning it to its original position. The elderly person is then asked to actively reproduce the angle with their eyes closed. The actual angle is recorded, and the single-joint deviation ratio is calculated by subtracting 1 from the target angle and dividing the difference between the actual and target angles by the absolute value of the target angle. The elbow and knee joint deviation ratios are calculated separately using the above method and then weighted and summed, with each weight being 0.5, to obtain the proprioceptive positioning accuracy. Visual and auditory parameters are normalized and mapped to the 0-100 range. A judgment matrix is ​​constructed: 3-5 rehabilitation medicine experts are invited to compare the importance of vision, hearing, and proprioception pairwise, and the weights of each parameter are obtained through matrix consistency testing. , and The corresponding values ​​are 0.4, 0.4 and 0.2 respectively.

[0120] In one embodiment of the present invention, step S5 includes the following steps:

[0121] The physiological health coefficient, cognitive health coefficient, and sensory health coefficient of the elderly are used as input vectors. ,in, The physiological health index of the elderly. The cognitive health index of the elderly Sensory health coefficient for the elderly;

[0122] The weights of each dimension are dynamically allocated using a neural attention mechanism.

[0123] The time-accumulated health data sequence is subjected to memory enhancement processing through a gated recurrent unit, and the output hidden state and the Hadamard product of the current input are used as feature interaction terms.

[0124] Random forest and XGBoost models are run in parallel, and the Bhattacharyya distance between the two models' predicted probability distributions is taken as the final confidence score.

[0125] The weighted composite index is normalized to the [0, 100] interval using the Sigmoid function to obtain the final comprehensive health score.

[0126] Specifically, the three health coefficients calculated in the early stage are used as model input to form an input vector. Initial weights are assigned to the three dimensions as the starting point of the attention mechanism. An additive attention model is used to generate dynamic weights, and attention scores are calculated. After a linear transformation of the input vector, it is activated by tanh, then normalized by softmax, and combined with a historical contribution decay factor to obtain the final weights of the corresponding dimensions. The historical health data sequence is then used to generate the dynamic weights. , , ..., ,in, At the current moment, the input is fed into a gated recurrent unit (GRU) to extract temporal dependency features. The current input vector is then multiplied by the Hadamard product of the temporal hidden state output by the GRU to obtain an interaction term that integrates historical and current information. The prediction process for the model includes: inputting the interaction term into a trained random forest model to output a health score probability distribution; inputting the interaction term into an XGBoost model to output a probability distribution; calculating the similarity distance between the two distributions, selecting the model output with the smallest distance as the confidence result; for the model output that passes the validation, a composite index is calculated using attention weights, and the composite index is mapped to the [0, 100] interval to obtain the final comprehensive health score. Specifically:

[0127]

[0128] in, For the final overall health score, This is the scaling factor. This is a weighted composite index. This is a midpoint offset to ensure that the overall health score is more sensitive in the middle range.

[0129] In one embodiment of the present invention, the dynamic allocation of weights for each dimension using a neural attention mechanism includes the following steps:

[0130] The weights of each dimension satisfy... ,and ,in, The historical contribution attenuation factor for the corresponding dimension;

[0131] The dynamic update process of the historical contribution decay factor is as follows:

[0132] For each health dimension Set the initial decay factor and define the forgetting factor;

[0133] After each backpropagation, calculate the partial derivative of the loss function with respect to the input features;

[0134] The decay factors for each dimension are refreshed iteratively based on exponential smoothing, specifically as follows:

[0135]

[0136] in, The updated decay factor, The historical contribution decay factor, i.e., the record cutoff time. At that time, the first The historical cumulative importance score of each feature Forgetting Gating Coefficient For the current loss function on the th Partial derivatives of each feature;

[0137] After all updates After normalization using the Softmax function, the dynamic weights for the next training cycle are obtained;

[0138] If a certain dimension in three consecutive iterations ,in, If the preset threshold is used, the learning rate for that dimension will be forcibly reset to the default value.

[0139] Specifically, set an initial decay factor for each dimension. , =1, 2, and 3 correspond to the physiological, cognitive, and sensory dimensions, respectively. The initial values ​​for each are set to one-third, evenly distributed, as the starting point for iteration. The forgetting factor is set to 0.1 to control the decay rate of historical data; a smaller value indicates faster forgetting of historical information. The prediction error of the comprehensive health score is used as the loss function, specifically: ,in, For sample size, For sample index, The overall health score output by the model. The reference score is based on expert evaluation; during each backpropagation of the model, the partial derivative of the loss function with respect to each input feature (i.e., the health coefficients of the three dimensions) is calculated, and the decay factor is iteratively updated using exponential smoothing based on the partial derivative results; if a certain dimension's health coefficients are low in three consecutive iterations... ,in, If the preset threshold is 0.05, the learning rate for that dimension will be forcibly reset to the default value of 0.01 to prevent the weights from reaching zero; if If the value is greater than 1, then it will be forced to be 0.99 to avoid excessive concentration of weight.

[0140] In one embodiment of the present invention, step S6, classifying health levels, includes the following steps:

[0141] The health level is divided into five levels based on the overall health score, as follows:

[0142] A health score of ≥85 is considered excellent; 70≤health score≤84 is considered good; 55≤health score≤69 is considered average; 40≤health score≤54 is considered poor; and a health score <40 is considered very poor.

[0143] The risk threshold was determined based on longitudinal tracking data of people in the same age group. The survival curve was fitted using the Weibull distribution, and the 10th percentile was used as the high-risk threshold.

[0144] Each level corresponds to a differentiated management protocol, including monitoring frequency, medication reminders, and emergency contact linkage mechanisms. Specifically, longitudinal tracking data of elderly people in the same age group is collected, including at least 3 years of comprehensive health score sequences and health outcomes, with a sample size of no less than 500 cases per stratum. The data is cleaned, and samples with a missing rate >20% are removed to obtain a valid dataset. The Weibull distribution model is used to fit the survival probabilities corresponding to different health scores, and the scale and shape parameters are solved by the maximum likelihood estimation method. The 10th percentile value of the comprehensive health score of the same age group is calculated as the high-risk threshold, i.e., comprehensive health score <40, ensuring that the survival probability at this threshold is ≤0.05, i.e., the risk of adverse outcomes is ≥50%. The remaining thresholds are adjusted by equal interval division combined with clinical experience: based on the 10th percentile, the division is carried out upwards at 15 percentile intervals to ensure that the survival probability gradient corresponding to each level is reasonable.

[0145] Please see Figure 2 As shown, the present invention is a question-and-answer type elderly health management method device, comprising:

[0146] Dynamic question answering module: Constructs a cognitive characteristic database of the elderly population and a dynamic question answering strategy library, which includes an initial question sequence and question adjustment rules;

[0147] Rhythm-aware scheduling module: Based on the rhythm analysis results in the cognitive characteristic database of the elderly population, prioritize the collection of health data of the elderly during the peak period of their daily activity.

[0148] Multimodal data acquisition module: Simultaneously collects physiological, cognitive, and sensory parameters of the elderly;

[0149] Health coefficient calculation module: For the three dimensions of physiology, cognition and senses, it generates the corresponding health coefficients for each dimension through interval mapping, standardized testing and normalization, including the physiological health coefficient, cognitive health coefficient and sensory health coefficient of the elderly.

[0150] Adaptive fusion computing module: It uses the neural attention mechanism to adaptively allocate cross-dimensional weights, inputs the historical accumulated health data sequence into the gated recurrent unit, extracts the time-dependent features and fuses them with the current interaction item, performs confidence verification through a dual-path machine learning model, and finally outputs a normalized comprehensive health score.

[0151] Health level assessment module: Analyzes the complete health information obtained, establishes an elderly health indicator system, performs Z-score standardization, uses the model to output the comprehensive health score of the elderly, and classifies the health level;

[0152] Closed-loop feedback optimization module: The health assessment report is fed back to the elderly and their relevant guardians, and the dynamic question-and-answer strategy library is updated based on the feedback information.

[0153] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A question-and-answer-based method for elderly health management, characterized in that, Includes the following steps: S1: Based on the cognitive characteristics database of the elderly population, a dynamic question-answering strategy library is constructed, which includes an initial question sequence and question adjustment rules; S2: Based on the rhythm analysis results in the cognitive characteristics database of the elderly population, health data collection for the elderly should be initiated during the peak periods of their daily activity levels. S3: Simultaneously collect physiological, cognitive, and sensory parameters of the elderly; S4: For the three dimensions of physiology, cognition and senses, the corresponding health coefficients for each dimension are generated through interval mapping, standardized testing and normalization, including the physiological health coefficient, cognitive health coefficient and sensory health coefficient of the elderly. S5: Adaptively allocate cross-dimensional weights using the neural attention mechanism, input the historically accumulated health data sequence into the gated recurrent unit, extract temporal dependent features and fuse them with the current interaction item, perform confidence verification through a dual-path machine learning model, and finally output a normalized comprehensive health score. S6: Analyze the complete health information obtained, establish an elderly health indicator system, perform Z-score standardization, use the model to output the comprehensive health score of the elderly, and classify the health level; S7: Provide health assessment reports to the elderly and their guardians, and update the dynamic question-and-answer strategy database based on the feedback information; Step S5 includes the following steps: The physiological health coefficient, cognitive health coefficient, and sensory health coefficient of the elderly are used as input vectors. ,in, The physiological health index of the elderly. The cognitive health index of the elderly Sensory health coefficient for the elderly; The weights of each dimension are dynamically allocated using a neural attention mechanism. The time-accumulated health data sequence is subjected to memory enhancement processing through a gated recurrent unit, and the output hidden state and the Hadamard product of the current input are used as feature interaction terms. Random forest and XGBoost models are run in parallel, and the Bhattacharyya distance between the two models' predicted probability distributions is taken as the final confidence score. The weighted composite index is normalized to the [0, 100] interval using the Sigmoid function to obtain the final comprehensive health score.

2. The question-and-answer-based elderly health management method according to claim 1, characterized in that, Step S1 includes the following steps: The elderly population cognitive characteristics database includes the elderly's memory level, language comprehension ability, reaction speed, and seasonal illness attack patterns. The question adjustment rules are formulated based on the semantic similarity of the answer information, the answer duration, the health language and number comprehension deviation index, and the diurnal rhythm characteristics. The circadian rhythm characteristics were determined using time series analysis to calculate the daily peak periods, trough periods, and other time periods for the elderly.

3. The question-and-answer-based elderly health management method according to claim 1, characterized in that, Step S2 includes the following steps: Based on the circadian rhythm characteristics in the cognitive characteristic database of the elderly population, the data collection work was started when the daily activity level of the elderly was at its peak. During the data collection process, elderly health data from various dimensions are recorded in real time, including physiological parameters, cognitive parameters, and sensory parameters. The physiological parameters include blood pressure, blood sugar, heart rate, and respiratory rate. The physiological health coefficient of the elderly is calculated based on the physiological parameters. The cognitive parameters include memory level, language comprehension ability, reaction speed, and health language and math comprehension deviation coefficient. The cognitive health coefficient of the elderly is calculated based on the cognitive parameters. The sensory parameters include visual parameters, auditory parameters, and proprioceptive positioning accuracy. The sensory health coefficient of the elderly is calculated based on the sensory parameters.

4. The question-and-answer-based elderly health management method according to claim 3, characterized in that, The physiological health coefficient of the elderly includes the following steps: For four physiological parameters, age-stratified health reference intervals were established for the elderly. The measured values ​​were converted into corresponding physiological health indices by interval mapping. Blood pressure was calculated by standard deviation weighting, while the other parameters were directly mapped. According to the preset weight allocation design, the four physiological health indices are weighted and summed. By introducing a metabolic age compensation factor to correct the weighted results, the physiological health coefficient of the elderly is obtained. The health reference interval for age stratification is determined based on the percentile of the elderly population in the same age group; the standard deviation weighting uses the coefficient of variation as the weighting coefficient; the metabolic age compensation factor is calculated by the ratio of actual physiological age measured by biomarkers to calendar age.

5. The question-and-answer-based elderly health management method according to claim 3, characterized in that, The cognitive health coefficient of the elderly includes the following steps: The standardized value of delayed recall test score was used as the memory level, the accuracy of semantic matching task was used as the language comprehension ability, the reaction speed was converted into the reaction efficiency value, which was obtained by dividing the actual reaction time by the age-matched benchmark reaction time, and the health language and math comprehension deviation index was determined by pre-testing to determine the cognitive error of professional vocabulary among the elderly. The corresponding weights are determined by combining the Delphi method with feature importance analysis using machine learning, and then weighted summation is performed to obtain the cognitive health coefficient of the elderly. Specifically: ； in, The cognitive health index of the elderly For memory level, For language comprehension ability, For reaction rate, For the health-related language and math comprehension deviation index, , , and These are the corresponding weights.

6. The question-and-answer-based elderly health management method according to claim 3, characterized in that, The sensory health coefficient of the elderly includes the following steps: Visual parameters, auditory parameters, and proprioceptive positioning accuracy parameters are obtained. The proprioceptive positioning accuracy is the ratio of the actual deviation angle to the theoretical value when an elderly person completes a joint angle reproduction task with their eyes closed. The visual and auditory parameters were normalized separately. The corresponding weights were determined using the analytic hierarchy process (AHP), and a weighted fusion model was constructed to obtain the sensory health coefficient of the elderly, specifically: ； in, The sensory health index of the elderly. and These are the normalized visual and auditory parameters, respectively. This refers to the proprioceptive positioning accuracy parameter. , and These are the corresponding weights.

7. The question-and-answer-based elderly health management method according to claim 1, characterized in that, The method of dynamically allocating weights for each dimension using a neural attention mechanism includes the following steps: The weights of each dimension satisfy... ,and ,in, The historical contribution attenuation factor for the corresponding dimension; The dynamic update process of the historical contribution decay factor is as follows: For each health dimension Set the initial decay factor and define the forgetting factor; After each backpropagation, calculate the partial derivative of the loss function with respect to the input features; The decay factors for each dimension are refreshed iteratively based on exponential smoothing, specifically as follows: ； in, The updated decay factor, The historical contribution decay factor, i.e., the record cutoff time. At that time, the first The historical cumulative importance score of each feature Forgetting Gating Coefficient For the current loss function on the th Partial derivatives of each feature; After all updates After normalization using the Softmax function, the dynamic weights for the next training cycle are obtained. If a certain dimension in three consecutive iterations ,in, If the preset threshold is used, the learning rate for that dimension will be forcibly reset to the default value.

8. The question-and-answer-based elderly health management method according to claim 1, characterized in that, In step S6, classifying health levels includes the following steps: The health level is divided into five levels based on the overall health score, as follows: A health score of ≥85 is considered excellent; 70≤health score≤84 is considered good; 55≤health score≤69 is considered average; 40≤health score≤54 is considered poor; and a health score <40 is considered very poor. The level threshold is determined based on longitudinal tracking data of people in the same age group, and the survival curve is fitted with Weibull distribution, with the 10th percentile as the high-risk threshold. Each level corresponds to a different management agreement, including monitoring frequency, medication reminders, and emergency contact linkage mechanisms.

9. A question-and-answer type elderly health management device, used to implement the question-and-answer type elderly health management method according to any one of claims 1-8, characterized in that, include: Dynamic question answering module: Based on the cognitive characteristics database of the elderly population, a dynamic question answering strategy library is constructed, which includes an initial question sequence and question adjustment rules; Rhythm-aware scheduling module: Based on the rhythm analysis results in the cognitive characteristic database of the elderly population, prioritize the collection of health data of the elderly during the peak period of their daily activity. Multimodal data acquisition module: Simultaneously collects physiological, cognitive, and sensory parameters of the elderly; Health coefficient calculation module: For the three dimensions of physiology, cognition and senses, it generates the corresponding health coefficients for each dimension through interval mapping, standardized testing and normalization, including the physiological health coefficient, cognitive health coefficient and sensory health coefficient of the elderly. Adaptive fusion computing module: It uses the neural attention mechanism to adaptively allocate cross-dimensional weights, inputs the historical accumulated health data sequence into the gated recurrent unit, extracts the time-dependent features and fuses them with the current interaction item, performs confidence verification through a dual-path machine learning model, and finally outputs a normalized comprehensive health score. Health level assessment module: Analyzes the complete health information obtained, establishes an elderly health indicator system, performs Z-score standardization, uses the model to output the comprehensive health score of the elderly, and classifies the health level; Closed-loop feedback optimization module: The health assessment report is fed back to the elderly and their relevant guardians, and the dynamic question-and-answer strategy library is updated based on the feedback information.

Images