Health status dynamic evaluation system and method for home-based elderly people based on multi-source information fusion

Abstract

The application discloses a home-based elderly health state dynamic evaluation system and method based on multi-source information fusion, which comprises the following steps: S1, collecting multi-source heterogeneous data covering the environment, physiology and behavior dimensions, and adopting a differentiated filling strategy for different types of multi-source heterogeneous data; S2, constructing an improved fuzzy comprehensive evaluation model, and outputting fuzzy evaluation features including index membership, fuzzy comprehensive evaluation results and dynamic health score sequences; S3, using an MCL-Informer model to perform multi-step prediction on future health scores, and outputting health score prediction values; S4, acquiring actual collection data at a time corresponding to the health score prediction values, obtaining actual dynamic health scores at the corresponding time by means of the improved fuzzy comprehensive evaluation model, detecting a prediction residual sequence, establishing a three-level early warning rule, and outputting corresponding level early warning signals according to a judgment condition. The application realizes whole-process closed-loop management from multi-source data collection, dynamic fuzzy evaluation, time series prediction to abnormal early warning.

Description

Technical Field

[0001] This invention relates to the fields of smart health and elderly care and intelligent information processing technology, specifically to a dynamic assessment system and method for the health status of home-based elderly people based on multi-source information fusion. Background Technology

[0002] Chronic disease health management in the home environment faces serious challenges: First, monitoring methods are often limited, typically focusing on isolated physiological indicators and failing to effectively integrate multi-source information such as environmental parameters (e.g., temperature and humidity) and individual behavioral data (e.g., activity level and sleep quality), leading to incomplete health status assessments. Second, existing assessment methods are mostly static and based on fixed thresholds, unable to adapt to dynamic changes in individual activity levels; for example, a normal heart rate at rest may be misjudged as abnormal during exercise. Third, there is a lack of effective predictive and early warning capabilities, making it difficult to predict health risk trends from massive amounts of continuous data, typically resulting in a reactive response only after symptoms appear.

[0003] In existing intelligent approaches, fuzzy comprehensive evaluation is used to handle the uncertainty of health data. However, its membership functions and indicator weights are often subjectively set or rely on simple data dispersion, failing to fully integrate medical logic (such as the weakest link effect) and easily getting trapped in local optima. Current time series predictions mostly employ LSTM and Transformer models. The former is effective for short-term fluctuations but lacks sufficient modeling of long-distance dependencies; the latter can capture global correlations but has limited response to local mutations and non-stationary fluctuations. Patents with publication numbers CN110866628A and CN115267575A, while introducing BiLSTM multi-scale convolution and Transformer prediction methods respectively, are still primarily driven by quantitative data and lack effective integration of qualitative knowledge such as expert experience and grade evaluation. In terms of anomaly early warning, the fixed threshold method has poor universality, and a single statistical method cannot simultaneously address sudden anomalies and gradual deterioration, making it difficult to balance overall detection rate and accuracy. Therefore, a comprehensive technical solution is urgently needed that can deeply integrate multi-source data, dynamically adapt to individual states, accurately predict future trends, and provide timely graded early warnings. Summary of the Invention

[0004] In order to overcome the above technical problems, the purpose of this invention is to provide a dynamic assessment system and method for the health status of home-based elderly people based on multi-source information fusion, so as to realize closed-loop management of the whole process from multi-source data collection, dynamic fuzzy assessment, time series prediction to abnormal early warning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A dynamic assessment method for the health status of elderly people living at home based on multi-source information fusion includes the following steps: S1: Collect multi-source heterogeneous data covering environmental, physiological, and behavioral dimensions. Environmental data includes temperature and humidity, physiological data includes heart rate, and behavioral data includes steps, activity minutes, and sleep scores. Differentiated imputation strategies are adopted for different types of multi-source heterogeneous data (such as cubic spline interpolation for continuous physiological signals to ensure their smoothness, and mean imputation for behavioral data) to form a high-quality, time-aligned multidimensional time-series dataset. S2: Based on the multidimensional time series dataset, construct an improved fuzzy comprehensive evaluation model and output fuzzy evaluation features including the membership degree of each indicator, fuzzy comprehensive evaluation results, and dynamic health score sequence. S3: Based on the multidimensional time-series dataset and the dynamic health score sequence, the MCL-Informer model is used to make multi-step predictions of future health scores and output the predicted health score values. S4: Constructing a dual-path adaptive early warning mechanism based on prediction residuals: Obtain the actual collected data at the time corresponding to the predicted health score, and input the actual collected data into the improved fuzzy comprehensive evaluation model in S2 to obtain the actual dynamic health score at the corresponding time; define the prediction residual as the difference between the actual dynamic health score and the predicted health score output in S3; Based on the predicted residual sequence, on the one hand, AR whitening combined with a nonparametric percentile method is used to set a dynamic threshold to identify point anomalies where residuals suddenly increase and waveform deviations of the actual score curve relative to the predicted score curve. On the other hand, the CUSUM cumulative offset detection method is used to accumulate and statistically analyze continuous residuals to identify the hidden trend of deterioration reflected by a health score that is continuously lower than the predicted level. Finally, a three-level early warning rule is established by combining the two types of detection results, and the corresponding level of early warning signal is output according to the judgment conditions. The method for multi-source heterogeneous data acquisition in step S1 is as follows: Using an indoor temperature and humidity recorder (such as T&D TR72A2-S) and a smart wearable bracelet (such as a device integrating heart rate and acceleration sensors), a chronic disease patient living at home was monitored synchronously for 7 consecutive days. The sampling frequency was once per minute, followed by uniform resampling at 5-minute intervals. The collected parameters included: ambient temperature (X1), ambient humidity (X2), resting heart rate (X3), heart rate (X4), sleep score (X5), and activity steps (X6). For missing values ​​that occurred during the collection process, such as heart rate data, cubic spline interpolation was used to impute them to maintain the continuous and smooth characteristics of the physiological signals. For step count data, mean imputation was used. After processing, a complete and aligned multidimensional time-series dataset was obtained. Time alignment, outlier removal, and missing value imputation are performed on the collected heterogeneous data from multiple sources, including environmental, physiological, and behavioral data.

[0006] In step S2, the multidimensional time-series dataset is used as the data basis for evaluation and weight optimization. Environmental, physiological, and behavioral indicators are extracted and their membership degrees are calculated. Step frequency and activity minutes are used as inputs, and the heart rate membership degree is dynamically corrected through a bivariate membership degree dynamic adjustment mechanism. At the same time, the GS-SSA algorithm, which integrates the golden sine and simplex reflection, is used to globally optimize the weights of each indicator, so that the weight allocation is adapted to the characteristics of multi-source data. The fitness function takes into account both the weakest link effect and the stability of the score, and finally outputs fuzzy evaluation features including the membership degrees of each indicator, fuzzy comprehensive evaluation results, and dynamic health score sequence.

[0007] The bivariate membership dynamic adjustment mechanism uses step frequency and activity minutes as input variables, where step frequency is calculated as follows: (1) in Effective activity cadence, steps / min; The cumulative number of steps taken within a 5-minute sliding window; The cumulative activity minutes (min) are used as the dual continuous inputs of step frequency and activity minutes. The activity status is divided into three categories: resting, light activity, and moderate to high intensity through a bivariate membership degree dynamic adjustment mechanism. 100 steps / minute is used as the threshold for moderate intensity physical activity, and an adaptive trapezoidal membership degree evaluation interval is configured for each category. Other indicators use a fixed trapezoidal membership degree function to map the measured values ​​at each time point to four levels: "excellent", "good", "average", and "poor". The GS-SSA algorithm is used to optimize the index weights of the improved fuzzy comprehensive evaluation model in S2. Based on the multidimensional time series dataset formed in S1, environmental, physiological and behavioral indicators are extracted and membership degrees are calculated. Among them, the heart rate membership degree is first corrected by a bivariate membership degree dynamic adjustment mechanism to match the activity status reflected by step frequency and activity minutes. On this basis, the weights of each evaluation index are used as the variables to be optimized in the GS-SSA algorithm, and the fitness function is constructed with the corrected membership degree matrix and health score results. The GS-SSA algorithm optimizes producer position updates by introducing the Golden Sine Operator (GSO), guiding candidate weight combinations toward the optimal solution. Simultaneously, it introduces a simplex adaptive reflection mechanism (SAR), calculating the centroid based on the top 10% elite individuals in the population and adjusting the reflection and shrinkage of the bottom 30% of individuals, thereby obtaining the optimal index weights that are compatible with the characteristics of multidimensional time-series data and the results of dynamic membership adjustment. The fitness function, which balances the weakest link effect and score stationarity, is defined as: (2) in This is the indicator weight vector; This is a comprehensive scoring time series; This represents the worst-performing indicator score sequence within each time window, reflecting health shortcomings. This is the balance coefficient; The two are related by Pearson; The optimal index weights obtained by optimizing the GS-SSA algorithm. A continuous health comprehensive score time series is obtained by fuzzy synthesis of the membership degree matrix at each time point. The optimal index weights are obtained by the GS-SSA algorithm, which uses the weakest link effect and score stability as fitness objectives, and performs global optimization on the weights of each evaluation index. The membership degree matrix is ​​calculated from the multidimensional time series dataset formed by S1, specifically including: normalizing the ambient temperature X1, ambient humidity X2, resting heart rate X3, heart rate X4, and sleep score X5, and converting them into corresponding health status membership degrees according to a preset membership degree function; the membership degree of heart rate X4 is further corrected by combining step frequency and activity minutes through a bivariate membership degree dynamic adjustment mechanism to match the heart rate evaluation result with the current activity state. Finally, the fuzzy synthesis result is mapped to a health comprehensive score, forming a dynamic health comprehensive score time series. In S3, the dynamic comprehensive health score time series output by S2 is used. To predict the target, a three-channel input structure is constructed: The MCL-Informer model first constructs three-channel input features: The first channel contains six raw observation features: temperature, humidity, heart rate, steps, activity minutes, and sleep score. The second channel is used to input the fuzzy evaluation results and dynamic health score sequence generated by the improved fuzzy comprehensive evaluation model in S2; The third channel consists of low-frequency steady-state derived features calculated based on the original observation features of S1 and the dynamic health score sequence of S2. The low-frequency steady-state derived features include the EWMA smoothing value of the health score, the long-term variance of heart rate, the cumulative value of environmental heat load, and the cumulative activity intensity. The three-channel features are independently embedded and encoded, then concatenated into a unified sequence, which is simultaneously input into a parallel dual-branch architecture of LSTM and Informer: the LSTM branch captures local short-term fluctuations and nonlinear changes through a gating mechanism, while the Informer branch efficiently extracts global long-range dependencies through a probabilistic sparse self-attention mechanism; the output features of the two branches are concatenated and then deeply fused by a fully connected layer to finally output a health score prediction for multiple future time steps. Original monitoring feature channel (containing six indicators from X1 to X6), fuzzy evaluation feature channel (containing The low-frequency steady-state characteristic channel includes the membership scores of each indicator, the 12-hour moving variance of heart rate, the cumulative characteristics of environmental heat load, and the cumulative number of steps in 30 minutes. Among them, the original monitoring characteristic channel includes six indicators X1 to X6 collected and preprocessed by S1, the fuzzy evaluation characteristic channel includes the membership scores of each indicator calculated by S2, and the low-frequency steady-state characteristic channel is obtained by further statistical calculation of the original monitoring data of S1 and the dynamic health comprehensive score time series of S2, including the EWMA smoothing characteristics of the comprehensive score, the 12-hour moving variance of heart rate, the cumulative characteristics of environmental heat load, and the cumulative number of steps in 30 minutes.

[0008] The three-channel features are independently embedded and encoded, and then concatenated to generate a time-series feature sequence. The input is a two-branch architecture that uses LSTM and Informer in parallel. The LSTM branch captures local short-term fluctuations and nonlinear changes through a gating mechanism, and outputs locally refined features. The Informer branch extracts global long-range dependencies using probabilistic sparse self-attention and hierarchical distillation mechanisms, outputting global evolutionary features. : (3) (4) In the formula, It is a long short-term memory network; This is a long sequence model based on a probabilistic sparse self-attention mechanism; The output features of the two branches are concatenated and then fused by a fully connected layer to output the future... Health score prediction at each time step In S4, the prediction residual at time t is defined as the difference between the actual health score and the predicted value: (5) when A value greater than 0 indicates that one's health condition is worse than expected and that there is a risk. The predicted residual sequence is subjected to first-order autoregressive whitening to reduce the temporal correlation between the predicted residuals, resulting in a whitened residual sequence; wherein, the predicted residual is the difference between the actual dynamic health score at the corresponding time and the predicted health score output by S3. (6) In the formula, For the intercept term, These are the autoregressive coefficients. The whitening residuals are based on historical whitening residuals, and the 95th percentile is used. and the 99th percentile As a dynamic threshold; At the same time, for situations where health conditions are worse than expected, [measures will be taken to ensure compliance with regulations]. Positive residuals >0 are accumulated and calculated to track small but persistent health status shifts: (7) In the formula, The standard deviation of the residual sequence. =0.5 is a reference value. =4.77 is the decision interval; By combining the detection results of AR whitening percentile and CUSUM methods, a three-level early warning rule is established: and This is the normal state; or A Level I warning is triggered at this time; or A Level II warning is triggered at this time; and The highest level III warning is triggered when the residual is negative; no warning is triggered when the corresponding condition is met for two consecutive time steps; a warning signal is officially output when the corresponding condition is met for two consecutive time steps. In this mechanism, AR whitening-percentile method is used to sensitively capture sudden point anomalies and waveform distortions, while CUSUM cumulative offset detection is used to identify hidden trend deterioration. The two are combined to achieve complementary detection and accurate classification of various anomalies.

[0009] The home-based elderly care health status dynamic assessment system based on multi-source information fusion includes a multi-source data acquisition module, a data preprocessing module, a dynamic fuzzy evaluation module, a health score prediction module, an adaptive early warning module, and a result output module. The system comprises the following modules: a multi-source data acquisition module for collecting environmental, physiological, and behavioral data; a data preprocessing module for performing time alignment, outlier removal, and missing value imputation on the collected data to form a multidimensional time-series dataset; a dynamic fuzzy evaluation module for calculating membership degrees, optimizing indicator weights, and outputting a dynamic comprehensive health score time series based on the multidimensional time-series dataset; a health score prediction module for predicting future health scores based on the multidimensional time-series dataset and the dynamic comprehensive health score time series; an adaptive early warning module for anomaly identification and graded early warning based on the prediction residual between the actual health score and the predicted health score; and a result output module for outputting the health score, prediction results, and early warning signals.

[0010] The modules are connected in sequence to form a closed-loop evaluation process of "data acquisition - data processing - health assessment - trend prediction - anomaly warning - result output".

[0011] The beneficial effects of this invention are: This invention employs a bivariate membership degree dynamic adjustment mechanism, using step frequency and activity minutes to determine the current activity state, and dynamically adjusts the membership degree evaluation interval of the heart rate index accordingly. This ensures that the heart rate evaluation results match resting, light activity, or moderate-to-high intensity activity states. After obtaining the adjusted membership degree matrices for each index, the GS-SSA algorithm is further used to globally optimize the index weights, and the optimal weight combination is selected through a fitness function that balances the weakest link effect and score stability. Thus, the evaluation criteria for the heart rate index are first adaptively adjusted according to the activity state, and the weights of each index are then objectively optimized based on the adjusted evaluation results. This overcomes the shortcomings of traditional methods, such as static thresholds being detached from the actual activity state, and the weights being highly subjective and prone to polarization, making the health score more closely reflect the individual's real-time physiological condition.

[0012] The MCL-Informer model of this invention uses a three-channel input structure to combine raw monitoring data, fuzzy evaluation results, and low-frequency steady-state derived features as prediction inputs, achieving deep fusion of qualitative assessment knowledge and quantitative monitoring data. Structurally, the three-channel features are embedded and encoded before being simultaneously input into an LSTM and an Informer. The LSTM branch uses gating units to retain state change information between adjacent time steps, capturing local short-term fluctuations and nonlinear changes in indicators such as heart rate, step count, and health score. The Informer branch uses a probabilistic sparse self-attention mechanism and a hierarchical distillation structure to extract global dependencies and trend changes over a longer time range. The fused features output predicted health scores for multiple future time steps, thus balancing local abnormal fluctuations and long-term trends, effectively mitigating the error accumulation problem in long-step prediction, and improving the accuracy and stability of multi-step prediction.

[0013] This invention employs a dual-path adaptive early warning mechanism based on predicted residuals. Specifically, a predicted residual sequence is first calculated based on the actual dynamic health score obtained in S2 and the predicted health score output in S3. On this basis, one path uses the AR whitening-nonparametric percentile method to decorrelate the predicted residual sequence and set a dynamic threshold to identify sudden anomalies reflected by short-term spikes in residuals and waveform deviations of the actual score curve relative to the predicted curve. The other path, for cases where the actual dynamic health score is consistently lower than the predicted score, uses CUSUM cumulative offset detection to accumulate and statistically analyze continuous small deviations, identifying a persistent but insignificant trend of declining health status. These two paths correspond to "sudden anomaly detection" and "trend deterioration detection," respectively. Finally, a three-level graded early warning is achieved through a joint judgment rule, effectively suppressing false alarms while maintaining a high detection rate.

[0014] The input data used in this invention can be obtained through conventional environmental sensors and commercial wearable devices. The calculation process is clear and applicable to home-based elderly people of different ages and with different types of chronic diseases. It has good cross-individual generalization ability and practical deployment feasibility. Attached Figure Description

[0015] Figure 1 This is a hierarchical logic block diagram for the activity state of bivariate dynamic fuzzy inference.

[0016] Figure 2 The graph shows the results of abnormal detection and graded early warning for 60M subjects (respondent IDs).

[0017] Figure 3 The graph shows the results of 60F abnormality detection and graded early warning for the subjects.

[0018] Figure 4 The figure shows the results of 61F abnormality detection and graded early warning in the subjects.

[0019] Figure 5 The image shows the results of abnormal detection and graded early warning for the subjects at 74M.

[0020] Figure 6 The figure shows the results of 85F abnormality detection and graded early warning for the subjects. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, this invention discloses a dynamic assessment system and method for the health status of home-based elderly people based on multi-source information fusion. In this embodiment, five home-based elderly subjects aged 60 and above with chronic diseases were selected for continuous multi-source monitoring and closed-loop verification. However, the types of subjects, scale, and monitoring period covered by this invention are not limited to this.

[0023] S1. Collect multi-source heterogeneous data covering environmental, physiological, and behavioral dimensions to construct an aligned multidimensional time-series dataset: In this case, five elderly subjects aged 60 and above with chronic diseases living at home were selected as monitoring subjects. An indoor temperature and humidity recorder (T&D TR72A2-S) was placed in the center of the subjects' living room (approximately 1.5m above the ground). Each subject also wore a smart wearable bracelet (integrating a photoelectric heart rate sensor and a three-axis accelerometer) for continuous synchronous monitoring for seven days. All sensors sampled at a uniform frequency of 1 time / minute, accumulating time-series data for six raw indicators: ambient temperature (X1), ambient humidity (X2), resting heart rate (X3), instantaneous heart rate (X4), sleep score (X5), and activity steps (X6).

[0024] After monitoring, the original 1-minute sampling sequences were aggregated and resampled in 5-minute windows, resulting in an aligned time-series dataset of 288 time steps per day for each subject, totaling 2016 time steps over 7 days. For missing values ​​caused by momentary removal of wearable devices, signal interruption, or abnormal battery levels during monitoring, differential imputation was performed based on the physical characteristics of each indicator: heart rate sequences X3 and X4 were imputed using cubic spline interpolation to maintain the continuity and smoothness of physiological signals; activity step count X6 was imputed using the mean within the corresponding 5-minute window to avoid misjudging activity intensity due to momentary missing values. The preprocessed multidimensional time-series datasets from the five subjects were merged and used as standardized input for subsequent S2 fuzzy comprehensive evaluation and S3 multi-step prediction.

[0025] S2. By constructing a bivariate dynamic membership mechanism and GS-SSA weighting, a dynamic comprehensive health score is output to evaluate the impact of each indicator on health status: First, a bivariate dynamic membership mechanism is constructed to address the disconnect between static thresholds and actual activity levels. Step frequency is calculated using formula (1). With step frequency and effective activity minutes as input, a fuzzy inference engine is used to classify human activity levels into three categories in real time: resting, light activity, and moderate to high intensity. Based on this classification result, the corresponding pre-set heart rate membership evaluation interval is invoked, as shown in Table 1. This achieves dynamic and adaptive adjustment of the heart rate index membership. Other indicators (temperature, humidity, resting heart rate, sleep score, and activity steps) use a fixed trapezoidal membership function, as shown in Table 2, mapping their measured values ​​to four levels of membership: "excellent," "good," "moderate," and "poor."

[0026] Table 1. Heart Rate Membership Criteria under Different Activity States state Excellent (bpm) Good(bpm) (bpm) Poor (bpm) resting state 60-70 70-80 50-60 / 80-90 <50 / >90 Light activity 70-90 90-110 60-70 / 110-130 <60 / >130 Medium to high strength 90-120 120-140 80-90 / 140-160 <80 / >160 Table 2 Evaluation Indicators: Membership Standards An improved sparrow search algorithm (GS-SSA) was used to optimize the index weights. A sparrow population was initialized, with each individual sparrow representing a weight vector of six indicators. In each iteration, the golden sine operator guided the discoverer to conduct a global exploration, and the simplex reflection mechanism was used to adjust individuals with low fitness, ensuring that the sum of the weights was 1 and continuously approaching the optimal solution. The fitness function was calculated according to formula (2), maximizing the correlation between the overall score and the weakest indicator, and constraining score fluctuations. After the iteration terminated, a set of globally optimal weight vectors was obtained, corresponding to X1~X6 respectively.

[0027] Finally, the optimized weights and the membership vectors at each time point are fuzzy synthesized, and then weighted and scored with the evaluation set "Excellent (100), Good (85), Average (70), Poor (55)" to finally output a continuous dynamic health comprehensive score time series.

[0028] S3. By constructing an MCL-Informer model for multi-step prediction, the prediction performance of different model and feature combinations is compared to verify the model's effectiveness: This step uses the comprehensive health score output by S2 as the prediction target and constructs a three-channel input structure: the original monitoring feature channel includes six indicators: temperature, humidity, resting heart rate, heart rate, sleep score, and activity steps; the fuzzy evaluation feature channel includes the dynamic fuzzy comprehensive score and the membership scores of each indicator, realizing the quantitative injection of qualitative knowledge; the low-frequency steady-state feature channel includes the EWMA smoothing value of the comprehensive score, the 12-hour moving variance of heart rate, the cumulative value of environmental heat load, and the cumulative activity intensity over 30 minutes. The qualitative domain knowledge of fuzzy evaluation is embedded into the quantitative prediction framework in the form of features, while low-frequency trend features reflecting the body's long-term steady-state regulatory capacity and cumulative environmental exposure are extracted to overcome the random noise interference of the original physiological data.

[0029] The three-channel features are independently embedded and encoded, then concatenated into a unified sequence, which is simultaneously input into a parallel dual-branch architecture of LSTM and Informer. The LSTM branch captures local short-term fluctuations and nonlinear changes through a gating mechanism according to formula (3), while the Informer branch extracts global long-range dependencies through probabilistic sparse self-attention and hierarchical distillation mechanisms according to formula (4). The outputs of the two branches are fused through a fully connected layer to output multi-step predictions. This allows the model to take into account both local details and global trends, effectively mitigating error accumulation in long-step predictions, and significantly improving prediction accuracy compared to a single-branch model.

[0030] S4. Construct a dual-path adaptive early warning mechanism based on prediction residuals to achieve hierarchical identification and early warning of abnormal health status: This step uses the combined abnormal injection data of 5 subjects as a basis to construct a dual-path adaptive early warning mechanism using the MCL-Informer health score prediction sequence output by S3. First, the difference between the actual health score and the predicted health score is calculated step by step according to formula (5) to obtain the prediction residual. This is used to characterize the degree of deviation of health status from the predicted baseline; then, the predicted residuals are analyzed according to formula (6). First-order autoregressive whitening was performed to obtain the whitening residual. This is used to respond to point anomalies and waveform anomalies; at the same time, the CUSUM positive cumulative statistic is calculated according to formula (7) to track continuous positive residual offsets and respond to trend anomalies. Based on this, the AR whitening-percentile method and the CUSUM cumulative offset detection results are combined to establish a three-level early warning rule of Level I (green), Level II (yellow), and Level III (red), and the early warning results are plotted as Figures 2-6 .Depend on Figures 2-6As can be seen, the gray curve represents the predicted residuals of the MCL-Informer model, the blue dashed line, the orange shading, and the red shading correspond to point anomalies, waveform anomalies, and trend anomalies, respectively, and the green, yellow, and red scattered points correspond to Level I, Level II, and Level III warnings, respectively. The results show that although there are individual differences in the baseline level and fluctuation amplitude of the residuals among the subjects, the warning scattered points are generally concentrated near the anomaly areas, and the warning level increases progressively with the degree of residual offset. Specifically, in the point anomaly area, a sudden increase in residual value occurs due to instantaneous pulse disturbances, triggering a Level III warning for most subjects; in the waveform anomaly area, the residuals exhibit continuous oscillations, with Level I to Level III warnings alternating, and the level changes are basically consistent with the strength of the fluctuation amplitude; in the trend anomaly area, the warning level gradually escalates with the accumulation of offset, initially dominated by Level I and Level II warnings, and later with a dense appearance of Level III warnings. These results indicate that this dual-path adaptive warning mechanism can identify point anomalies, waveform anomalies, and trend anomalies, and achieve a graded response to the risk of health deterioration.

Claims

1. A method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous data covering environmental, physiological, and behavioral dimensions. Environmental data in the multi-source heterogeneous data includes temperature and humidity, physiological data includes heart rate, and behavioral data includes steps, activity minutes, and sleep scores. For different types of multi-source heterogeneous data, a differential imputation strategy is adopted to form a multi-dimensional time-series dataset. S2: Based on the multidimensional time series dataset, construct an improved fuzzy comprehensive evaluation model and output fuzzy evaluation features including the membership degree of each indicator, fuzzy comprehensive evaluation results, and dynamic health score sequence. S3: Based on the multidimensional time-series dataset and the dynamic health score sequence, the MCL-Informer model is used to make multi-step predictions of future health scores and output the predicted health score values. S4: Obtain multi-source heterogeneous data corresponding to the predicted health score at the time, and input the actual collected data into the improved fuzzy comprehensive evaluation model in S2 to obtain the actual dynamic health score at the corresponding time; define the prediction residual as the difference between the actual dynamic health score and the predicted health score output in S3; Based on the predicted residuals, an AR whitening method combined with a nonparametric percentile method is used to set a dynamic threshold to identify point anomalies where the residuals suddenly increase and waveform deviations of the actual score curve relative to the predicted score curve. The CUSUM cumulative offset detection method is used to accumulate and statistically analyze continuous residuals to identify the hidden trend of deterioration reflected by the health score being continuously lower than the predicted level. Finally, a three-level early warning rule is established by combining the two types of detection results, and the corresponding level of early warning signal is output according to the judgment conditions.

2. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 1, characterized in that, The method for acquiring multi-source heterogeneous data in step S1 is as follows: Using an indoor temperature and humidity recorder and a smart wearable bracelet, a chronic disease patient living at home was continuously monitored. The sampling frequency was once per minute, followed by uniform resampling at 5-minute intervals. The parameters collected included: ambient temperature X1, ambient humidity X2, resting heart rate X3, heart rate X4, sleep score X5, and activity steps X6. Time alignment, outlier removal, and missing value imputation are performed on the collected heterogeneous data from multiple sources, including environmental, physiological, and behavioral data.

3. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 1, characterized in that, In step S2, the multidimensional time-series dataset is used as the data basis for evaluation and weight optimization. Environmental, physiological and behavioral indicators are extracted and their membership degrees are calculated. Using step frequency and activity minutes as inputs, the heart rate membership is dynamically corrected through a bivariate membership dynamic adjustment mechanism. At the same time, the GS-SSA algorithm, which integrates the golden sine and simplex reflection, is used to globally optimize the weights of each indicator, so that the weight allocation is adapted to the characteristics of multi-source data. The final output includes fuzzy evaluation features including the membership of each indicator, fuzzy comprehensive evaluation results, and dynamic health score sequence.

4. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 3, characterized in that, The bivariate membership dynamic adjustment mechanism uses step frequency and activity minutes as input variables, where step frequency is calculated as follows: (1) in Effective activity cadence, steps / min; The cumulative number of steps taken within a 5-minute sliding window; The cumulative activity minutes are min; with step frequency and activity minutes as dual continuous inputs, the activity state is divided into three categories: resting, light activity, and moderate to high intensity through a bivariate membership dynamic adjustment mechanism. 100 steps / minute is used as the threshold for moderate intensity physical activity, and an adaptive trapezoidal membership evaluation interval is configured for each category. Other indicators use a fixed trapezoidal membership function to map the measured values ​​at each time point to four levels: "excellent, good, average, and poor". The GS-SSA algorithm is used to optimize the index weights of the improved fuzzy comprehensive evaluation model in S2. Based on the multidimensional time series dataset formed in S1, environmental, physiological and behavioral indicators are extracted and membership degrees are calculated. Among them, the heart rate membership degree is first corrected by a bivariate membership degree dynamic adjustment mechanism to match the activity status reflected by step frequency and activity minutes. On this basis, the weights of each evaluation index are used as the variables to be optimized in the GS-SSA algorithm, and the fitness function is constructed with the corrected membership degree matrix and health score results.

5. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 4, characterized in that, The GS-SSA algorithm optimizes producer position updates by introducing the golden sine operator, guiding candidate weight combinations toward the optimal solution. At the same time, it introduces a simplex adaptive reflection mechanism, calculates the centroid based on the top 10% elite individuals in the population, and performs reflection and shrinkage adjustments on the bottom 30% of individuals, thereby obtaining the optimal index weights that are adapted to the characteristics of multidimensional time series data and the results of dynamic membership adjustment. The fitness function, which balances the weakest link effect and score stationarity, is defined as: (2) in This is the indicator weight vector; This is a comprehensive scoring time series; This represents the worst-performing indicator score sequence within each time window, reflecting health shortcomings. This is the balance coefficient; The two are related by Pearson; The optimal index weights obtained by optimizing the GS-SSA algorithm. Fuzzy synthesis was performed with the membership matrix at each time point to obtain a continuous health comprehensive score time series; The optimal index weights are the weight combinations obtained by the GS-SSA algorithm after global optimization of the weights of each evaluation index with the short-board effect and score stability as fitness objectives. The membership matrix is ​​calculated from the multidimensional time-series dataset formed by S1. Specifically, it includes: normalizing the ambient temperature X1, ambient humidity X2, resting heart rate X3, heart rate X4, and sleep score X5, and converting them into corresponding health status memberships according to a preset membership function; wherein, the membership of heart rate X4 is further modified by combining step frequency and activity minutes through a bivariate membership dynamic adjustment mechanism to match the heart rate evaluation result with the current activity status; finally, the fuzzy synthesis result is mapped to a comprehensive health score to form a dynamic comprehensive health score time series. .

6. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 5, characterized in that, In S3, the dynamic comprehensive health score time series output by S2 is used. To predict the target, a three-channel input structure is constructed: The MCL-Informer model first constructs three-channel input features: The first channel contains six raw observation features: temperature, humidity, heart rate, steps, activity minutes, and sleep score. The second channel is used to input the fuzzy evaluation results and dynamic health score sequence generated by the improved fuzzy comprehensive evaluation model in S2; The third channel consists of low-frequency steady-state derived features calculated based on the original observation features and the dynamic health score sequence of S2. The low-frequency steady-state derived features include the EWMA smoothing value of the health score, the long-term variance of the heart rate, the cumulative value of the environmental heat load, and the cumulative activity intensity. The three-channel features are independently embedded and encoded, then concatenated into a unified sequence, which is synchronously input into a parallel dual-branch architecture of LSTM and Informer: the LSTM branch captures local short-term fluctuations and nonlinear changes through a gating mechanism, while the Informer branch efficiently extracts global long-range dependencies through a probabilistic sparse self-attention mechanism; the output features of the two branches are concatenated and then deeply fused by a fully connected layer to finally output a health score prediction for multiple future time steps.

7. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 6, characterized in that, The three-channel features are independently embedded and encoded, and then concatenated to generate a time-series feature sequence. The input is a two-branch architecture that uses LSTM and Informer in parallel. The LSTM branch captures local short-term fluctuations and nonlinear changes through a gating mechanism, and outputs locally refined features. The Informer branch extracts global long-range dependencies using probabilistic sparse self-attention and hierarchical distillation mechanisms, outputting global evolutionary features. : (3) (4) In the formula, It is a long short-term memory network; This is a long sequence model based on a probabilistic sparse self-attention mechanism; The output features of the two branches are concatenated and then fused by a fully connected layer to output the future... Health score prediction at each time step .

8. The method for dynamic assessment of the health status of home-based elderly people based on multi-source information fusion according to claim 7, characterized in that, In S4, the prediction residual at time t is defined as the difference between the actual health score and the predicted value: (5) when A value greater than 0 indicates that one's health condition is worse than expected and that there is a risk. The predicted residual sequence is subjected to first-order autoregressive whitening to obtain a whitened residual sequence; wherein, the predicted residual is the difference between the actual dynamic health score at the corresponding time and the predicted health score output by S3: (6) In the formula, For the intercept term, These are the autoregressive coefficients. The whitening residuals are based on historical whitening residuals, and the 95th percentile is used. and the 99th percentile As a dynamic threshold; At the same time, for situations where health conditions are worse than expected, [measures will be taken to ensure compliance with regulations]. Positive residuals >0 are accumulated and calculated to track small but persistent health status shifts: (7) In the formula, The standard deviation of the residual sequence; By combining the detection results of AR whitening percentile and CUSUM methods, a three-level early warning rule is established: and This is the normal state; or A Level I warning is triggered at this time; or A Level II warning is triggered at this time; and The highest level (Level III) warning is triggered when the residual is negative; no warning is triggered when the corresponding conditions are met for two consecutive time steps. The warning signal is officially output when the corresponding conditions are met for two consecutive time steps.

9. A dynamic assessment system for the health status of home-based elderly people based on multi-source information fusion, used to implement the method described in any one of claims 1-8, characterized in that, It includes a multi-source data acquisition module, a data preprocessing module, a dynamic fuzzy evaluation module, a health score prediction module, an adaptive early warning module, and a result output module; The system comprises the following modules: a multi-source data acquisition module for collecting environmental, physiological, and behavioral data; a data preprocessing module for performing time alignment, outlier removal, and missing value imputation on the collected data to form a multidimensional time-series dataset; a dynamic fuzzy evaluation module for calculating membership degrees, optimizing indicator weights, and outputting a dynamic comprehensive health score time series based on the multidimensional time-series dataset; a health score prediction module for predicting future health scores based on the multidimensional time-series dataset and the dynamic comprehensive health score time series; an adaptive early warning module for anomaly identification and graded early warning based on the prediction residual between the actual health score and the predicted health score; and a result output module for outputting the health score, prediction results, and early warning signals.

Images