Elderly chronic disease discovery and management system

Abstract

The invention relates to the technical field of senile chronic disease management, and discloses a discovery and management system based on senile chronic diseases. The system comprises a health data acquisition module, a risk sensitivity assessment module, a disease risk positioning module, a management instruction generation module and a health report generation module. The health data acquisition module obtains real-time physiological signals through a wearable sensor, and precise multi-scale health features are obtained through layered feature extraction; and the risk sensitivity evaluation module calculates a deviation degree, recognizes an abnormal region and fuses environmental data to generate a sensitivity index. The disease risk positioning module classifies diseases according to indexes, determines high-risk nodes and extracts time sequence data; and the management instruction generation module analyzes the key event and outputs a personalized executable instruction. And the health report generation module receives the feedback data, evaluates the index trend and generates a dynamic report. The system realizes early discovery, fine evaluation and strong intervention of chronic diseases, and improves the management efficiency and the health guarantee level of patients.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of elderly chronic disease management, in particular to a system for discovering and managing elderly chronic diseases. BACKGROUND

[0002] The incidence of elderly chronic diseases is on the rise, and chronic diseases such as hypertension, diabetes, and coronary heart disease have become major factors threatening the health of the elderly. Such diseases have the characteristics of long course, hidden progression, and high risk of complications, which have a serious impact on the quality of life of patients and also bring a heavy medical burden to families and society. Therefore, early detection, accurate assessment, and scientific management of elderly chronic diseases have become an important issue in the current medical and health field.

[0003] Currently, the management of elderly chronic diseases relies mainly on traditional regular physical examinations and patient self-reported health data. This mode has obvious limitations. Regular physical examinations often have long intervals, making it difficult to capture real-time changes in the development of the disease, and the warning ability for sudden illness or short-term abnormalities is insufficient, which may miss the best intervention opportunity. Patient self-reported data is limited by the cognitive level and operational ability of the elderly, and the accuracy, completeness, and timeliness of the data cannot be guaranteed. Some patients may even miss key health information due to negligence or complacency, making it difficult for medical personnel to fully understand the patient's true health status.

[0004] With the development of sensing technology and information technology, some health management systems have begun to introduce wearable devices for data collection, but existing systems still have many defects. In terms of data processing, most systems only extract simple numerical values from physiological signals, failing to deeply mine the multi-scale features contained in the signals, resulting in low utilization efficiency of health data and difficulty in reflecting the complexity of the patient's health status. In the risk assessment stage, existing technologies often only focus on the deviation of physiological indicators from the baseline value, ignoring the spatial dependence between deviations, and failing to accurately locate the abnormal health pattern area. Moreover, the consideration of environmental factors is insufficient, while temperature, humidity, air quality, and other environmental factors often have a significant impact on the condition of elderly patients with chronic diseases, and the lack of assessment in this dimension will lead to deviations in risk judgment.

[0005] The existing system also has obvious short boards in the pertinence and dynamics of disease management. Most systems provide general health management programs, which fail to develop personalized plans according to the individual health characteristics and risk sensitivity differences of patients, resulting in a significant reduction in the effectiveness of management measures. At the same time, the health management process lacks a closed-loop mechanism. After issuing management instructions, the system cannot collect health feedback data from patients in a timely manner, nor can it dynamically adjust management strategies based on changes in the disease, making it difficult to achieve continuous and precise control of the disease. The existence of these problems makes the existing elderly chronic disease management system difficult to meet the actual clinical needs, and a new management system that can achieve accurate data collection, comprehensive risk assessment, personalized and dynamic management is needed. SUMMARY

[0006] The purpose of the present application is to provide an elderly chronic disease discovery and management system based on the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides an elderly chronic disease discovery and management system based on the problems raised in the background art. A health data acquisition module for continuously acquiring real-time physiological signals of elderly chronic disease patients through wearable sensors and performing hierarchical feature extraction on the real-time physiological signals to obtain multi-scale health feature representations; A risk sensitivity evaluation module based on the multi-scale health feature representations, calculates the deviation degree from the baseline health pattern, analyzes the spatial dependence relationship in the deviation degree, identifies abnormal health pattern areas, and simultaneously fuses environmental factor data to calculate the sensitivity index of the abnormal health pattern areas to environmental changes; A disease risk positioning module based on the sensitivity index, classifies the abnormal health pattern areas by disease type, generates a disease risk probability distribution, and locates high-risk nodes as weak links from the health management network graph, performs real-time health monitoring on the weak links, and extracts health status time series data; A management instruction generation module that analyzes key health events in the health status time series data, generates a personalized health management plan, and optimizes the health management plan for constraints to output an executable health management instruction set; A health report generation module that collects health feedback data after implementing the health management instruction set, evaluates the stability trend of health indicators, and generates a dynamic health report based on the stability trend.

[0008] Preferably, the health data acquisition module performs hierarchical feature extraction on the real-time physiological signals to obtain multi-scale health feature representations, including: Dividing the real-time physiological signals into overlapping time windows; applying a restricted Boltzmann machine to the data in each time window to obtain a local feature vector; inputting the local feature vectors of multiple time windows into a stacked autoencoder to perform global feature fusion and generate a multi-scale health feature representation.

[0009] Preferably, the risk sensitivity evaluation module calculates a deviation degree from a benchmark health pattern based on the multi-scale health feature representation, including: loading the benchmark health pattern from a health knowledge base as a reference feature set; calculating Mahalanobis distance between the multi-scale health feature representation and the reference feature set; performing exponential smoothing on the Mahalanobis distance to obtain a time-weighted deviation degree.

[0010] Preferably, the analysis of the spatial dependence in the deviation degree identifies an abnormal health pattern region, including: constructing a health feature graph, where nodes represent health feature dimensions and edges represent correlations between features; performing node embedding learning on the health feature graph using a graph convolution network to obtain a feature embedding vector; identifying a node set deviating from the normal distribution in the feature embedding space as an abnormal health pattern region through an anomaly detection algorithm.

[0011] Preferably, the risk sensitivity evaluation module uses a multivariate regression model to calculate a sensitivity index of the abnormal health pattern region to environmental changes, including: collecting environmental factor data, including temperature, humidity, and air quality index; establishing a multivariate regression equation with environmental factors as independent variables and feature values of the abnormal health pattern region as dependent variables; fitting the regression coefficients by least squares method, and taking the sum of absolute values of the coefficients as the sensitivity index.

[0012] Preferably, the disease risk positioning module locates a high-risk node as a weak link from a health management network graph using a community detection algorithm, including: constructing a health management network graph, where nodes represent health management components and edges represent interaction relationships between components; applying Louvain algorithm to community division of the health management network graph to obtain community structure; calculating the modularity of each community, and identifying the node in the community with the lowest modularity as a high-risk node.

[0013] Preferably, the disease risk positioning module performs real-time health monitoring on the weak link to extract health state time series data, including: Biological sensors are deployed at weak links to collect physiological parameters at a fixed frequency; The collected physiological parameters are subjected to wavelet transform to extract time-frequency features; The time-frequency features are arranged in chronological order to form health state time series data.

[0014] Preferably, the management instruction generation module generates a personalized health management plan using a sequence-to-sequence model, including: Encode the health state time series data into a hidden state sequence; Use an attention mechanism to weight and aggregate the hidden state sequence to generate a context vector; Convert the context vector into a health management plan sequence through a decoder.

[0015] Preferably, the management instruction generation module optimizes the health management plan with constraints, including: Retrieve constraint conditions of relevant treatment protocols from a clinical knowledge base; Establish a constraint satisfaction problem model, taking the health management plan as a variable domain; Use a backtracking algorithm to search for an optimized solution of the health management plan that satisfies all constraint conditions.

[0016] Preferably, the health report generation module uses a volatility analysis algorithm to evaluate the stability trend of health indicators, including: Calculate the first-order difference sequence of health feedback data; Perform autocorrelation analysis on the first-order difference sequence to detect periodic fluctuations; Based on the periodic fluctuations, use a sliding window to calculate local variance and fit a variance curve to evaluate the stability trend.

[0017] Compared with the prior art, the present application has the following advantages: In the health data collection link, the system discards the traditional single numerical extraction method and uses wearable sensors to continuously capture real-time physiological signals of elderly patients, ensuring the continuity and timeliness of data collection, effectively avoiding the time interval limitations of regular physical examinations and the lagging problem of self-reported data. On this basis, multi-scale health feature representations are obtained through hierarchical feature extraction techniques, which can analyze the health information contained in physiological signals from different dimensions and levels. Not only does it cover basic physiological indicator values, but also includes deep features such as signal change trends and fluctuation rules, fully releasing the value of health data and providing more comprehensive and accurate data support for subsequent risk assessment.

[0018] The design of the risk sensitivity evaluation module breaks through the evaluation limitations of the prior art. On the basis of calculating the deviation degree of physiological indicators from the benchmark health pattern, the spatial dependence relationship in the deviation degree is further analyzed, which can accurately identify the abnormal health pattern area, and changes the previous situation that only the overall index abnormality can be found but the specific problem link cannot be located. At the same time, the module integrates environmental factor data into the evaluation system, fully considers the influence of environmental variables such as temperature, humidity and air quality on the condition of elderly patients with chronic diseases, obtains the sensitivity index of the abnormal area to environmental changes through fusion analysis, and makes the risk evaluation result more suitable for the actual living scene of the patient, which provides an important basis for subsequent development of targeted management measures and effectively avoids the risk misjudgment caused by neglecting environmental factors.

[0019] The disease risk positioning module realizes fine management of disease risk. Based on the sensitivity index, the abnormal health pattern area is classified by disease type, which can determine the potential disease risk corresponding to different abnormal areas, and the generated disease risk probability distribution provides an intuitive risk reference for medical personnel. By positioning the high-risk nodes in the health management network diagram as weak links and performing real-time health monitoring on them, the key prevention and control of disease risk can be realized, and the limited health management resources can be concentrated on the key links. At the same time, the health state time series data of the weak links can be extracted to dynamically track the risk change trend and capture the subtle changes of the disease progression in time, which provides dynamic data support for the adjustment of subsequent management plan.

[0020] The management instruction generation module focuses on individualized health management needs. By analyzing the key health events in the health state time series data, the individual health characteristics, disease change law and risk sensitive points of the patient can be accurately grasped, and the individualized health management plan generated accordingly breaks the limitations of the existing general management scheme and fully meets the physical condition, living habit and disease characteristics of different patients. At the same time, by optimizing the health management plan, the actual constraint conditions such as the execution ability of the patient and the accessibility of medical resources can be considered comprehensively, so that the generated health management instruction set has good operability, avoids the problem that the previous management plan is too complex or unrealistic to be executed, improves the patient's compliance, and makes the health management measures truly effective.

[0021] The health report generation module constructs a closed-loop feedback mechanism for health management. After the implementation of management instructions, health feedback data is collected in a timely manner. By evaluating the stability trend of health indicators, the implementation effect of management measures can be objectively and comprehensively reflected. The dynamic health report generated based on the stability trend can not only enable patients and medical personnel to clearly grasp the condition change situation, but also provide a basis for the adjustment of subsequent management plans. When the health indicators show a stable and favorable trend, the current management strategy can be continued. When the indicators fluctuate or deteriorate, the reasons can be analyzed in a timely manner and the management measures can be optimized and adjusted to ensure that health management always adapts to the changes in the patient's condition, providing continuous and precise health protection for elderly patients with chronic diseases, effectively improving the patient's quality of life, and reducing the medical burden on families and society. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A working principle diagram of the system for discovering and managing elderly chronic diseases based on the application; Figure 2 A flowchart of the hierarchical feature extraction work of the health data collection module; Figure 3 A flowchart of the work of the risk sensitivity assessment module to identify abnormal health pattern areas; Figure 4 A time-weighted deviation degree change trend chart; Figure 5 A health management plan variable constraint satisfaction rate comparison chart. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the application.

[0024] Please refer to Figure 1The application provides a chronic disease discovery and management system for the elderly, which includes a health data acquisition module that collects real-time physiological signals of patients, such as electrocardiogram, blood oxygen, or activity data, through wearable devices such as heart rate monitors or blood pressure sensors; after preprocessing, these signals enter the hierarchical feature extraction stage, divide the time series into overlapping windows, and apply deep learning models to learn local and global features to form multi-scale health feature representations. The risk sensitivity evaluation module receives these feature representations, compares them with pre-stored baseline health patterns, calculates the deviation, and analyzes the spatial dependence between features to identify abnormal areas; this module integrates environmental factors such as temperature or humidity data, and quantifies the sensitivity of abnormal areas to environmental changes through regression analysis. The disease risk positioning module classifies abnormal areas based on the sensitivity index, generates a disease risk probability distribution, and locates high-risk nodes from the health management network graph as weak links; these nodes are monitored more closely, and time series data is extracted for further analysis. The management instruction generation module analyzes key events in the time series data, uses sequence models to generate personalized health management plans, and ensures the feasibility of the plans through constraint optimization, outputting an executable instruction set. The health report generation module collects feedback data after the implementation of the instructions, evaluates the stability trend of the health indicators, and generates dynamic reports to support long-term management.

[0025] Example 1: see Figure 2In specific implementations, the process of hierarchical feature extraction on real-time physiological signals to obtain multi-scale health feature representation by the health data acquisition module involves a series of rigorous data processing steps. Real-time physiological signals are continuously collected by wearable sensors, including but not limited to photoplethysmography, electrocardiogram, and body motion signals. The raw signals collected are converted into discrete time series data by an analog-to-digital converter. In specific implementations, the signal preprocessing stage includes power line notch filtering to remove power supply interference and baseline drift correction based on median filtering. The preprocessed signals are fed into the hierarchical feature extraction process. The core of hierarchical feature extraction is to divide the continuous real-time physiological signal stream into a series of overlapping time windows. The length of the time window is dynamically configured according to the characteristics of the physiological signal. For example, for electrocardiogram signals, the time window length is set to 5 seconds, and the overlap is set to 50% of the window length, i.e. 2.5 seconds. This overlapping division method ensures that transient features in the signal are not truncated by the window boundary. Restricted Boltzmann machines are applied to the discrete data within each time window for unsupervised feature learning. Restricted Boltzmann machines are a type of probabilistic graphical model that includes a visible layer and a hidden layer. The number of visible layer units is consistent with the number of sampling points in the time window, and the number of hidden layer units is empirically set to 128. The training process updates the connection weights between the visible layer and the hidden layer using the contrastive divergence algorithm, allowing the hidden layer to learn the underlying statistical regularities in the time window data. Finally, a fixed-dimensional local feature vector is output from the hidden layer of the restricted Boltzmann machine for each time window. The local feature vector encodes the local patterns of the signal in short time scales, such as the morphological features of QRS complexes or the slight variations in pulse intervals.

[0026] In specific implementations, the local feature vectors corresponding to multiple consecutive time windows are organized in chronological order into a sequence, which is used as input to a stacked autoencoder for global feature fusion. The stacked autoencoder is composed of multiple encoder layers and decoder layers. The encoder layers are usually composed of fully connected neural networks that map high-dimensional local feature vectors to low-dimensional representations through nonlinear activation functions. The decoder layers attempt to reconstruct the original input from the low-dimensional representation, with the training goal being to minimize the reconstruction error function. In some embodiments, the stacked autoencoder adopts a three-layer structure, with the first layer of encoder compressing the local feature vector from 128 dimensions to 64 dimensions, and the second layer further compressing to 32 dimensions. The third layer of decoder performs the reverse reconstruction process, and finally extracts a global feature vector that can represent long-term dependencies from the output of the bottleneck layer of the stacked autoencoder, i.e. the second layer of encoder. The multi-scale health feature representation is obtained by concatenating the local feature vector and the global feature vector. The local feature vector captures high-frequency details and short-term fluctuations in the signal, while the global feature vector captures macro trends and periodic patterns in the signal. This multi-scale representation provides a rich information foundation for subsequent risk assessment.

[0027] It can be understood that the risk sensitivity evaluation module calculates the deviation degree from the benchmark health pattern based on the multi-scale health feature representation, and the benchmark health pattern is a reference feature set loaded from a health knowledge base. The health knowledge base is a pre-constructed database storing a large amount of historical physiological data collected in a stable state of healthy elderly individuals and their extracted feature representations, and the benchmark health pattern is obtained by K-means clustering algorithm processing on these historical feature representations. The cluster center vector is used as the reference feature set. In specific implementation, the Mahalanobis distance between the multi-scale health feature representation and the reference feature set is calculated, and the calculation of the Mahalanobis distance requires the covariance matrix of the reference feature set, which describes the correlation structure between the dimensions of the health features. The Mahalanobis distance can effectively measure the deviation degree of the feature representation to be evaluated relative to the distribution of the healthy population, and its calculation process involves solving the inverse matrix of the covariance matrix and multiplying it with the quadratic form of the feature difference vector. The initial Mahalanobis distance value is a scalar reflecting the instantaneous deviation.

[0028] In specific implementation, the calculated Mahalanobis distance sequence is subjected to exponential smoothing to obtain a time-weighted deviation degree, and exponential smoothing is a time series smoothing technique, which basically assigns greater weight to recent observations. The smoothing process follows a recursive formula, in which the smoothed value at the current time is a weighted average of the original observation at the current time and the smoothed value at the previous time. The smoothing coefficient a ranges from 0 to 1, and the closer the value of a is to 1, the more sensitive it is to recent data. The time-weighted deviation degree is finally output as a continuous sequence of values, which not only reflects the instantaneous difference between the current features and the benchmark pattern, but also incorporates historical deviation information, thus being able to more stably represent the dynamic trend of the health status. In some embodiments, the smoothing coefficient a is set to 0.3 to achieve a balance between smoothing effect and response speed.

[0029] It can be understood that the quality of the multi-scale health feature representation directly affects the accuracy of the deviation degree calculation, so the selection of parameters in the hierarchical feature extraction process, such as the length of the time window, the overlap ratio, the number of hidden layer units of the restricted Boltzmann machine, and the structure of the stacked autoencoder, needs to be carefully adjusted according to the specific application scenario and signal type. The deviation degree, as the basis for subsequent sensitivity index calculation and abnormal pattern recognition, its reliability depends on the representativeness of the benchmark health pattern and the numerical stability of the Mahalanobis distance calculation. In specific implementation, the health knowledge base needs to be updated regularly to include new health individual data, ensuring that the benchmark health pattern can reflect the overall characteristics of the current population.

[0030] Embodiment 2: see Figure 3In specific implementations, the process of analyzing the spatial dependency in the bias degree and identifying the abnormal health pattern region by the risk sensitivity assessment module starts with the construction of a health feature graph. The health feature graph is a kind of graph structure data representation, in which the nodes correspond to the dimensions of the multi-scale health feature representation, for example, node one represents the average heart rate feature, node two represents the heart rate variability feature, and node three represents the blood oxygen saturation feature. The number of nodes is completely consistent with the dimension of the feature vector. The edges are used to represent the statistical correlation between different health feature dimensions, and the weight of the edge is determined by calculating the Pearson correlation coefficient between the data sequences of each two feature dimensions. The Pearson correlation coefficient measures the degree of linear correlation, and the greater the absolute value, the stronger the correlation between the feature dimensions. The constructed health feature graph contains the topological connection information of the health feature system.

[0031] In specific implementations, the constructed health feature graph is subjected to node embedding learning using a graph convolutional network. The graph convolutional network is a neural network specially used for processing graph structure data, and its core operation is to generate a new node representation by aggregating the feature information of each node and its neighbor nodes. Each layer of the graph convolutional network performs feature propagation and transformation once. For each node in the health feature graph, the graph convolutional network collects feature information from its directly connected neighbor nodes, combines the node's own features, performs linear transformation through a learnable weight matrix, and then passes through a nonlinear activation function such as the ReLU function to output the updated feature representation of the node at the current layer. After multiple layers of graph convolution operations, each health feature dimension node will obtain a low-dimensional feature embedding vector rich in graph structure information. The feature embedding vector can capture the relative position and local connection pattern of the feature dimension in the global graph.

[0032] In specific implementations, the set of nodes deviating from the normal distribution in the feature embedding space composed of feature embedding vectors is identified by an anomaly detection algorithm. The anomaly detection algorithm regards the set of vector points in the entire feature embedding space as a whole distribution, and the algorithm needs to load a "normal" feature embedding distribution model learned from the health individual data as a reference benchmark from the health knowledge base. In some embodiments, the anomaly detection algorithm used is the local outlier factor algorithm, which calculates a local outlier factor score for each feature embedding vector. This score measures the local density deviation of the vector point from its k nearest neighbor vector points. If the local density of a feature embedding vector point is significantly lower than that of its neighbor points, then this feature embedding vector point is determined to be an outlier. The original health feature dimension nodes corresponding to all feature embedding vector points determined to be outliers are collectively identified as an abnormal health pattern region. The abnormal health pattern region indicates that the cooperative change behavior of certain feature combinations has exceeded the normal range.

[0033] It can be understood that the risk sensitivity assessment module uses a multiple regression model to calculate the sensitivity index of the abnormal health pattern region to environmental changes is an independent but associated step. Collecting environmental factor data is a prerequisite for this step, including temperature, humidity and air quality index, which are obtained through an interface with environmental monitoring sensors or access to public environmental data services, with strictly aligned sampling timestamps and physiological signal collection timestamps. The characteristic value of the abnormal health pattern region usually refers to some kind of aggregate statistics of the multiple feature dimension vectors that make up the region, such as the first principal component score after principal component analysis, which is used as the dependent variable for multiple regression analysis.

[0034] In specific implementation, a multiple regression equation is established in the form of a linear additive model, with environmental factor data as independent variables and characteristic values of the abnormal health pattern region as dependent variables. The multiple regression equation contains an intercept term and a regression coefficient corresponding to each environmental factor independent variable. The regression coefficients are fitted by least squares method, which finds a set of regression coefficient estimates by solving a system of normal equations, so that the sum of the squares of the differences between the predicted values and the true values of the dependent variable for all observation data points reaches a minimum. The fitting process needs to ensure that there is no serious multicollinearity problem between independent variables, otherwise it will affect the stability of the regression coefficient estimates. After fitting, each environmental factor independent variable will get a regression coefficient, which quantifies the amount of change in the characteristic value of the abnormal health pattern region expected to occur when the environmental factor changes by one unit.

[0035] In specific implementation, the absolute values of the regression coefficients of each environmental factor independent variable obtained by fitting are summed, and the sum is the sensitivity index. The absolute value is taken because the sign of the regression coefficient only indicates the direction of the impact, while the sensitivity index needs to measure the size or strength of the impact. The sensitivity index is a comprehensive scalar index, and the higher the value, the more sensitive the overall state of the abnormal health pattern region to the overall fluctuations of the environmental factors. In some embodiments, if the multiple regression analysis includes temperature, humidity and air quality index as three independent variables, then the sensitivity index is equal to the sum of the absolute values of the temperature regression coefficient, the humidity regression coefficient and the air quality index regression coefficient. Optionally, before calculating the sensitivity index, the environmental factor data can be standardized to eliminate the influence of different environmental factor dimensions and numerical range differences on the absolute value of the regression coefficient, making the interpretation of the sensitivity index more consistent.

[0036] It can be understood that the number of layers of the graph convolution network, the selection of the activation function, the setting of the parameter k in the anomaly detection algorithm, and the screening of the independent variables in the multiple regression model are all configurable hyperparameters. The construction quality of the health feature map directly depends on the accuracy of the Pearson correlation coefficient calculation between features, and requires a long enough time span of clean data to ensure stable estimation of the correlation coefficient. The dimension of the feature embedding vector needs to be low enough to facilitate subsequent anomaly detection, but also needs to be high enough to retain the necessary distinguishing information. The identification of the abnormal health pattern area provides a clear analysis target for subsequent disease risk positioning, while the sensitivity index quantifies the potential impact of external environmental factors on health risks, both of which provide important data-driven insights for personalized health management. The effectiveness of the multiple regression model is based on the assumption that there is a linear relationship between environmental factors and health features. For nonlinear relationships, an optional extension is to introduce higher-order terms or interaction terms of environmental factors into the regression equation.

[0037] Referring to Figure 4 The chart belongs to the risk sensitivity assessment module of the elderly chronic disease discovery and management system, and is a visual presentation of the time-weighted deviation degree calculation logic. The horizontal axis is the monitoring time point, covering 60 consecutive monitoring nodes, representing the dynamic tracking process of the health status; the vertical axis is the deviation degree, quantifying the deviation of the patient's real-time health features from the baseline health pattern. This chart is one of the core bases for risk sensitivity assessment: the fluctuation of the original deviation degree can assist in identifying short-term abnormal events, and the trend of the time-weighted deviation degree provides basic data support for the identification of abnormal health pattern areas and the calculation of environmental change sensitivity indexes. Medical personnel can intuitively judge the persistence and severity of the patient's health deviation through this chart, timely capture potential chronic disease progression risks, and provide quantitative basis for the development of subsequent personalized intervention strategies.

[0038] In a specific implementation, the disease risk positioning module utilizes a community detection algorithm to locate high-risk nodes from the health management network graph as weak links, which starts with the construction of the health management network graph. The health management network graph is a directed or undirected graph structure, in which nodes are used to represent various health management components in the system, and specific examples of health management components include a medication reminder and dosage recording component, a personalized exercise plan generation component, a daily diet nutrition recommendation component, a physiological data trend analysis component, and a doctor-patient communication log component. Edges are used to represent the interaction or dependency relationships between health management components, including data flow relationships, control flow relationships, or function call relationships, for example, the medication reminder and dosage recording component sends a message to the personalized exercise plan generation component after adjusting the dosage, and there is a directed edge between the two components. The weight of the edge can be quantitatively assigned by the frequency of communication between components, the size of the data volume, or the coupling strength of the function. The completed health management network graph abstractly depicts the internal structure of the entire health management system and the dynamic collaboration pattern between components.

[0039] In a specific implementation, the Louvain algorithm is applied to the constructed health management network graph to obtain the community structure. The Louvain algorithm is a hierarchical clustering community detection algorithm based on modularity optimization, and the algorithm execution process is iterated in two stages. In the first stage, the algorithm initially assigns each node in the health management network graph to a community, then traverses each node in the network, attempts to move the node to the community where its neighbor nodes are located, calculates the change ΔQ in the modularity Q value before and after the movement, and the modularity Q value measures the difference between the internal connection density of the community and the expected density under random connection. If a community movement can bring positive modularity gain ΔQ, and it is the largest gain among all possible movements, then move the node to the corresponding community. The first stage will continue until no node movement can improve the overall modularity Q value. In the second stage, the Louvain algorithm contracts each community identified in the first stage into a new super node, the edge weight between super nodes is the sum of the weights of all edges between the original communities, and the internal edge weight of the community is converted to the self-loop weight of the new super node. Then, the algorithm repeats the first stage process based on the contracted new network graph. Such iteration continues until the network structure no longer changes and the modularity Q value reaches a local maximum, and finally outputs a hierarchical community structure, in which each community contains a set of tightly connected health management components.

[0040] In a specific implementation, the modularity of each community is calculated to identify weak links. The modularity contribution value of each community can be calculated by the following formula:

[0041] wherein the symbol Modularity contribution value of community c, denoted as Sum of weights of all edges within community c, denoted as Sum of weights of all edges in the whole health management network graph, denoted as Sum of weights of all edges connected to nodes in community c (including edges connecting within the community and edges connecting to outside communities). The overall modularity Q of the whole network is the sum of all community modularity contribution values .

[0042] After calculating the modularity contribution values of all communities, the community with the lowest modularity contribution value is identified as the relatively weak area of the system structure. It can be understood that the community with a low modularity contribution value means that the internal connections are too sparse relative to its connections with other parts of the network, and the component collaboration relationship within the community may not be close enough or there may be functional isolation. All health management component nodes in this community are identified as high-risk nodes, which represent potential weak links that may fail or have performance bottlenecks under system operating pressure. In some embodiments, a threshold value of the modularity contribution value can be set, and multiple communities with contribution values below the threshold value are considered as high-risk communities.

[0043] It can be understood that real-time health monitoring of the identified weak links and extraction of health status time series data are the basis for subsequent analysis. In specific implementation, deploying biosensors at weak links means strengthening the monitoring of relevant physiological parameters for the specific health management function represented by the identified high-risk nodes. For example, if the high-risk node is a "high-intensity exercise safety monitoring component", then an electrocardiogram sensor and an oxygen saturation sensor need to be deployed for intensive monitoring when the patient is exercising as planned. Biosensors collect physiological parameters at a fixed frequency, and the fixed frequency is set according to the characteristics of the parameters, such as a heart rate signal acquisition frequency of 250 Hz and an oxygen saturation frequency of 1 Hz. The collected raw physiological parameter signals need to be preprocessed, including denoising and normalization.

[0044] In a specific implementation, the pre-processed physiological parameters are subjected to wavelet transform to extract time-frequency features. Wavelet transform employs a basis function called mother wavelet to analyze the characteristics of a signal at different time and frequency scales through translation and scaling operations. Unlike Fourier transform that only provides frequency information, wavelet transform is capable of providing both time localization and frequency localization information, making it very suitable for analyzing non-stationary physiological signals. For a discrete physiological parameter sequence collected, discrete wavelet transform is applied to decompose the signal into different levels of approximation coefficients and detail coefficients using multi-resolution analysis. These coefficients collectively constitute the time-frequency feature representation of the signal, characterizing the energy intensity of a specific frequency component at a particular time instant. The time-frequency features extracted from wavelet transform are arranged in chronological order to form health status time series data. The time series data is a three-dimensional array structure, with the first dimension being the time point sequence, the second dimension being the feature index, and the third dimension being the feature value itself. The health status time series data systematically records the dynamic evolution process of the weak link related physiological state over time. Optionally, before forming the final health status time series data, the time-frequency features can be subjected to dimensionality reduction processing, such as using principal component analysis method to reduce the data volume and retain the main variation information. The health status time series data serves as the direct input of the management instruction generation module, providing a data basis for generating personalized health management plans. In some embodiments, the health status time series data will be stored and indexed in the form of a time series database for efficient query and retrospective analysis.

[0045] In a specific implementation, the management instruction generation module uses a sequence-to-sequence model to generate a personalized health management plan. The process starts with the encoding of health status time series data. Health status time series data is a multi-dimensional time series that contains the time-varying characteristics of time-frequency features extracted from weak link monitoring. The sequence-to-sequence model adopts an encoder-decoder architecture, with the encoder part responsible for mapping variable-length input sequences into a fixed-dimensional context vector representation. The encoder is usually composed of a recurrent neural network, such as a long short-term memory network or a gated recurrent unit, which can effectively capture long-term dependencies in time series. The health status time series data is input into the encoder long short-term memory network step by step, and the long short-term memory network updates its internal state at each time step based on the current input and hidden state. Finally, after processing the entire input sequence, the final hidden state sequence of the encoder long short-term memory network constitutes a compressed representation of the input health status time series data, i.e., the hidden state sequence. Each vector in the hidden state sequence contains the sequence information up to the current time step.

[0046] In a specific implementation, the hidden state sequence of the encoder is weightedly aggregated using an attention mechanism to generate a context vector. The attention mechanism allows the model to dynamically focus on the importance of different parts of the input sequence when generating each time step of the output sequence. The computation of the attention mechanism is as follows: for the hidden state of the decoder at a certain time step, the attention scores between it and each hidden state of the encoder are calculated, which are usually calculated by a trainable feedforward neural network, representing the degree of association between the current state of the decoder and each hidden state of the encoder. Then these attention scores are normalized by the Softmax function to obtain the attention weight distribution, which reflects the importance of each encoder hidden state to the current decoding step. Finally, the attention weights are used to weight the sum of all hidden states of the encoder to generate a context vector related to the current decoding step. The context vector is a fixed-length vector that condenses the most relevant information in the input sequence to the current output, providing a dynamic and focused input for the decoder to generate a personalized health management plan.

[0047] In a specific implementation, the context vector is converted into a specific health management plan sequence by the decoder. The decoder is another recurrent neural network, such as a long short-term memory network, which takes the context vector and the output of the previous time step as input to generate each element of the health management plan step by step. The health management plan sequence can be represented as a series of instruction tokens, such as "increase", "drug A", "dose", "10mg", "after meal", etc. The decoder long short-term memory network takes the context vector and the sequence start symbol as input at the initial time step to produce a probability distribution of the first output token, and selects the token with the highest probability as output through the Argmax operation or sampling. Then this output token is taken as the input of the next time step, while the current hidden state of the decoder long short-term memory network is passed to the next step, and this process is repeated until the sequence end symbol is generated or the preset maximum length is reached. The final generated token sequence constitutes a structured, personalized health management plan based on the individual's current health status, which may include drug adjustment recommendations, activity suggestions, dietary considerations, etc.

[0048] It is understood that after the preliminary health management plan is generated, the management instruction generation module needs to perform constraint optimization on it to ensure the safety and feasibility of the plan. The constraint optimization process retrieves the constraint conditions of the relevant treatment protocols from the clinical knowledge base, which is a structured database storing rules such as drug compatibility contraindications, dose ranges, administration times, patient-specific contraindications (such as dose adjustments for renal dysfunction), activity intensity restrictions, etc. The retrieval process is based on the elements involved in the health management plan, for example, if a certain drug is included in the plan, all constraint conditions related to that drug are retrieved. These constraint conditions form the boundaries of the optimization problem.

[0049] In specific implementation, a constraint satisfaction problem model is established to formalize the optimization problem of the health management plan. In the constraint satisfaction problem model, each adjustable element in the health management plan is modeled as a variable, such as a drug dose variable, a medication time variable, and a movement type variable. Each variable has a corresponding domain that defines all possible values of the variable, for example, the domain of the drug dose variable may be a set of discrete values between the minimum effective dose and the maximum safe dose. The constraint conditions retrieved from the clinical knowledge base are then converted into relationships that must be satisfied between variables, which define the legal combinations of variable values. For example, the constraint condition "drug A cannot be used simultaneously with drug B" is converted into a logical constraint that the "drug A use flag" and the "drug B use flag" cannot be true at the same time. The goal of the constraint satisfaction problem model is to find a set of variable assignments that satisfy all constraint conditions while being as close as possible to the initial health management plan generated by the sequence-to-sequence model.

[0050] In specific implementation, referring to Table 1, a backtracking algorithm is used to search for an optimized health management plan that satisfies all constraint conditions. Backtracking is a systematic search algorithm that traverses the variable assignment combination space in a depth-first manner. The algorithm starts with the first variable and assigns it a value from its domain, then checks if the constraints related to the assigned variables are satisfied. If satisfied, the algorithm continues to assign values to the next variable. If, while assigning a value to a variable, no value can be found that satisfies all current constraints, the algorithm backtracks to the previous variable and tries the next possible value for that variable. In this way, the backtracking algorithm can exhaust all possible variable assignment combinations until it finds a complete assignment that satisfies all constraint conditions, which is the optimized health management plan. If there are multiple feasible solutions, the one with the smallest difference from the initial plan can be selected as the final solution. The efficiency of the backtracking algorithm can be improved through forward checking, constraint propagation, and other strategies. The optimized health management plan ensures that its recommended content conforms to medical standards and the individual situation of the patient, and the final output is a clear and executable set of health management instructions.

[0051] Table 1: Health management plan variables and constraint conditions

[0052] Referring to Figure 5 The chart is one of the core outputs of the elderly chronic disease discovery and management system management instruction generation module, used to quantify the degree of fit of the personalized health management plan to the clinical constraints. The horizontal axis is the constraint satisfaction rate, reflecting the proportion of health management plan variables that meet the constraint conditions in the clinical knowledge base; the vertical axis is the health management plan variable, covering five types of core management elements: drug A dosage, exercise type, medication time, dietary sodium intake, and review frequency, which are directly related to key dimensions such as patient medication safety, exercise feasibility, and treatment timeliness. The business value of this chart is that it visually verifies the effectiveness of the system constraint satisfaction problem model + backtracking algorithm, retrieves constraint conditions from the clinical knowledge base, establishes a variable domain model, and systematically searches for feasible solutions, ultimately generating a health management plan that not only fits the individual health characteristics of the patient but also strictly adheres to medical standards. Medical personnel can quickly identify areas with relatively low constraint fit and optimize management strategies accordingly to ensure the executability and safety of the health management instruction set, thereby improving the precision and adherence of elderly chronic disease management.

[0053] In specific implementation, the health report generation module uses a fluctuation analysis algorithm to evaluate the stability trend of health indicators, which begins with preprocessing of health feedback data. Health feedback data is a series of time series indicators collected after implementing the health management instruction set, such as daily average blood pressure, blood glucose level, or activity duration, which are derived from continuous monitoring records of wearable devices or regular manual input by patients. The raw health feedback data collected may have missing values or obvious entry errors, so data cleaning is needed, including using linear interpolation to fill in missing values and using a quartile-based outlier detection method to remove erroneous data points that deviate significantly from the normal range, resulting in a complete and relatively clean health feedback data sequence for subsequent analysis.

[0054] In specific implementation, calculating the first-order difference sequence of health feedback data is a key step in fluctuation analysis. The first-order difference sequence is obtained by calculating the difference between adjacent time point observations in health feedback data. For a time-ordered health indicator sequence, the first-order difference value at the tth time point is equal to the tth observation minus the (t-1)th observation. The first-order difference operation effectively eliminates the trend component in the time series, shifting the focus of analysis from the absolute level of the indicator to its period-to-period change, which represents the short-term fluctuation of health. For example, for a 30-day continuous daily systolic blood pressure record, a first-order difference sequence containing 29 values is calculated, each representing the change in systolic blood pressure compared to the previous day, with positive values indicating an increase and negative values indicating a decrease.

[0055] In specific implementations, autocorrelation analysis is performed on the first-order difference sequence to detect periodic fluctuations. Autocorrelation analysis is used to quantify the linear correlation between the first-order difference sequence and its lagged versions, computing a series of autocorrelation coefficients at different lag orders. The autocorrelation coefficient ranges from -1 to 1, and the closer its absolute value is to 1, the stronger the periodicity. By examining the autocorrelation function plot or the autocorrelation coefficient values, one can identify whether there is a significant periodic pattern, e.g., a 7-day periodic fluctuation might correspond to the weekly lifestyle habits' impact on the health indicator. In some embodiments, the autocorrelation coefficients at lag 1 to lag 10 are focused on to capture the periodic patterns from short-term to medium-term. The identified periodic fluctuations provide important information for understanding the inherent rhythm of the health indicator changes.

[0056] In specific implementations, based on the identified periodic fluctuations, a sliding window is used to compute the local variance to characterize the variation in fluctuation intensity. A sliding window is a fixed-length subsequence that moves point by point along the time axis. For the first-order difference sequence, a window size is set, e.g., 7 days, and starting from the beginning of the sequence, the variance of all first-order difference values within the first window is computed, then the window is moved one time point forward, the variance of the next window is computed, and so on until the end of the sequence. Variance is a statistical measure of how much the data points deviate from their mean, and the larger the variance within a window, the more intense the short-term fluctuations of the health indicator in that time period. By computing the local variance for a series of consecutive windows, a local variance sequence is obtained, which describes the evolution of fluctuation intensity over time. Optionally, if significant periodicity is found by autocorrelation analysis, the window size can be set to the length of the period to better match the inherent rhythm of the fluctuations.

[0057] In specific implementations, a trend line is fitted to the local variance sequence to assess the stability trend. The local variance sequence itself is also a time series, and by fitting a trend line to it, one can visually judge the overall direction of fluctuation. The fitting process usually employs least squares method for polynomial regression or non-parametric smoothing methods like local weighted regression. For example, for a 90-day health management period, 84 local variance values with a window size of 7 days are computed, and a linear regression is fitted to these variance values. If the slope of the fitted line is negative, it indicates that the fluctuation of the health indicator is decreasing overall, meaning that the health status is tending to be stable; on the contrary, if the slope is positive, it means that the fluctuation is increasing, and the stability is decreasing. Optionally, instead of a linear trend, a quadratic curve can also be fitted to capture more complex trend variation patterns. Optionally, the determination coefficient of the trend line can also be computed to assess the significance of the trend.

[0058] It can be appreciated that the final generated dynamic health report will incorporate the above analysis results. The report will not only contain the trend plots of the original health indicators, but also highlight the key change points revealed by the first-order difference series, the periodic patterns identified by the autocorrelation analysis, and the stability conclusions indicated by the local variance trend curve. For example, the report can state: "In the past four weeks, the short-term volatility of blood pressure shows a significant downward trend, indicating the effectiveness of the intervention measures on stabilizing blood pressure." This data-driven stability assessment provides objective and quantitative feedback on the effectiveness for both doctors and patients, which helps guide the adjustment of subsequent health management strategies.

[0059] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences of the principles and spirit of the present application and that numerous modifications, changes, substitutions, and alterations can be made thereto without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A system for chronic disease discovery and management in the elderly, comprising: The system comprises: a health data acquisition module for continuously acquiring real-time physiological signals of elderly patients with chronic diseases through wearable sensors and performing hierarchical feature extraction on the real-time physiological signals to obtain multi-scale health feature representations; a risk sensitivity evaluation module for calculating a deviation degree from a baseline health pattern based on the multi-scale health feature representations, analyzing spatial dependence in the deviation degree, identifying abnormal health pattern regions, and calculating a sensitivity index of the abnormal health pattern regions to environmental changes by fusing environmental factor data; a disease risk positioning module for classifying disease types of abnormal health pattern regions based on the sensitivity index, generating a disease risk probability distribution, and positioning high-risk nodes as weak links from a health management network graph for real-time health monitoring of the weak links to extract health status time series data; a management instruction generation module for analyzing key health events in the health status time series data, generating a personalized health management plan, and performing constraint optimization on the health management plan to output an executable health management instruction set; a health report generation module for collecting health feedback data after implementing the health management instruction set, evaluating stability trends of health indicators, and generating a dynamic health report based on the stability trends.

2. The system of claim 1, wherein, The health data acquisition module performs hierarchical feature extraction on the real-time physiological signals to obtain multi-scale health feature representations, including: dividing the real-time physiological signals into overlapping time windows; applying a restricted Boltzmann machine to data in each time window for unsupervised feature learning to obtain local feature vectors; inputting local feature vectors of multiple time windows into a stacked autoencoder for global feature fusion to generate multi-scale health feature representations.

3. The system of claim 1, wherein, The risk sensitivity evaluation module calculates a deviation degree from a baseline health pattern based on the multi-scale health feature representations, including: loading a baseline health pattern from a health knowledge base as a reference feature set; calculating Mahalanobis distances between the multi-scale health feature representations and the reference feature set; performing exponential smoothing on the Mahalanobis distances to obtain time-weighted deviation degrees.

4. The system of claim 3, wherein, The analysis of spatial dependence in the deviation degree to identify abnormal health pattern regions includes: constructing a health feature graph, where nodes represent health feature dimensions and edges represent correlations between features; using a graph convolution network to perform node embedding learning on the health feature graph to obtain feature embedding vectors; identifying a node set deviating from a normal distribution in the feature embedding space as an abnormal health pattern region through an anomaly detection algorithm.

5. The system of claim 1, wherein, The risk sensitivity evaluation module calculates a sensitivity index of the abnormal health pattern regions to environmental changes using a multivariate regression model, including: collecting environmental factor data, including temperature, humidity, and air quality index; establishing a multivariate regression equation with environmental factors as independent variables and feature values of abnormal health pattern regions as dependent variables; fitting regression coefficients through least squares method, and taking the sum of absolute values of the coefficients as the sensitivity index.

6. The system of claim 1, wherein, The disease risk positioning module positions high-risk nodes as weak links from a health management network graph using a community detection algorithm, including: Construct a health management network graph, where nodes represent health management components and edges represent interaction relationships between components; Apply the Louvain algorithm to community division of the health management network graph to obtain community structure; Calculate the modularity of each community, and identify the nodes in the community with the lowest modularity as high-risk nodes.

7. The system of claim 1, wherein, The disease risk positioning module performs real-time health monitoring on the weak link, extracts health status time series data, including: Deploying a biosensor at the weak link to collect physiological parameters at a fixed frequency; Performing wavelet transform on the collected physiological parameters to extract time-frequency features; Arranging the time-frequency features in chronological order to form health status time series data.

8. The system of claim 1, wherein, The management instruction generation module generates a personalized health management plan using a sequence-to-sequence model, including: Encoding the health status time series data into a hidden state sequence; Using an attention mechanism to weight and aggregate the hidden state sequence to generate a context vector; Converting the context vector to a health management plan sequence through a decoder.

9. The system of claim 1, wherein, The management instruction generation module optimizes the health management plan by constraints, including: Retrieving constraint conditions of relevant treatment protocols from a clinical knowledge base; Establishing a constraint satisfaction problem model with the health management plan as the variable domain; Using a backtracking algorithm to search for an optimized solution of the health management plan that satisfies all constraint conditions.

10. The system of claim 1, wherein, The health report generation module uses a fluctuation analysis algorithm to evaluate the stability trend of health indicators, including: Calculating the first-order difference sequence of health feedback data; Performing autocorrelation analysis on the first-order difference sequence to detect periodic fluctuations; Based on the periodic fluctuations, using a sliding window to calculate the local variance and fitting the variance curve to evaluate the stability trend.

Images