Chronic disease risk automatic early warning method based on multi-modal data analysis

Abstract

The invention relates to a chronic disease risk automatic early warning method based on multi-modal data analysis, and belongs to the technical field of intelligent medical treatment and health management. The method comprises the following steps: acquiring multi-modal health data of a user in real time through a multi-modal data acquisition module, and performing feature extraction to obtain individual multi-dimensional health features; based on the features, a convolutional neural network and a graph neural network are called to be combined with the chronic disease type to construct a chronic disease risk prediction model, and potential chronic disease risks are judged through leaving-one verification; analyzing a risk level according to the risk result and generating early warning information; and carrying out risk detection based on the evaluation index, and if the risk is higher than a model set threshold, sending a risk early warning signal through the terminal device. Accurate assessment and personalized early warning of chronic disease risks are realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical and health management technology, specifically relating to an automatic early warning method for chronic disease risks based on multimodal data analysis. Background Technology

[0002] Chronic diseases, characterized by long course, insidious onset, and prolonged intervention periods, have become a major threat to global public health. Traditional chronic disease monitoring and management models rely heavily on regular hospital checkups, clinical consultations, and patient self-reporting, which suffer from problems such as delayed response, untimely risk detection, and neglect of individual differences. In recent years, with the widespread application of wearable devices, smartphones, and telemedicine platforms, users' health data can be continuously collected in non-medical environments. This data involves physiological parameters, behavioral habits, medical images, and medical records, exhibiting significant multimodal, heterogeneous, and high-frequency characteristics. This provides a data foundation for early risk identification of chronic diseases, but it also brings challenges to the complexity of analysis and processing. Existing chronic disease risk assessment methods are mostly based on single-modal data for modeling, which makes it difficult to fully capture the coupling relationship and potential correlation between different modalities, limiting the fine perception of complex health states. At the same time, although some studies have introduced artificial intelligence models, these models are mostly biased towards static prediction and lack dynamic tracking and individualized adjustment capabilities, making it difficult to adapt to changes in users' health status. Therefore, there is an urgent need for an intelligent chronic disease risk identification method that integrates multimodal data, has high-dimensional feature expression capabilities, and a dynamic early warning mechanism, so as to achieve early detection and early intervention of individual chronic disease risks and improve the scientific and intelligent level of chronic disease prevention and control. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, this invention provides an automatic early warning method for chronic disease risks based on multimodal data analysis. The objective of this invention can be achieved through the following technical solutions: S1: Collect users' multimodal health data in real time through the multimodal data acquisition module, and perform dimensionality reduction processing on the multimodal health data based on the principal component analysis method to construct an individual's multidimensional health set; S2: Based on the extracted multidimensional health set, the multimodal data is input into a recurrent feedforward network to extract spatial features. At the same time, the chronic disease record information is transformed into a graph structure of nodes and edges to train the chronic disease risk classification model. The learning rate of the parameters is adjusted by an adaptive matrix estimator. The collected multimodal health data is input into the chronic disease risk classification model. The average classification accuracy and absolute error are calculated based on leave-one-out validation to analyze potential chronic disease risks. S3: Based on the potential chronic disease risk, a risk level range is constructed by combining the individual's current physiological parameters and medical indicator change trends through a nearest neighbor feature embedding model, and multimodal risk intervention information is generated based on the risk level range. S4: The multimodal risk intervention information uses a multimodal risk convolution mechanism, combined with the risk level fluctuation trend, to determine the current risk status using multidimensional health features and the multimodal risk intervention information as an observation sequence. If the current risk status is detected to exceed the safe risk level range, a risk warning signal is automatically issued through the medical platform.

[0004] Specifically, the multimodal data acquisition module includes: a medical record access unit and a physiological signal acquisition unit. The medical record access unit accesses the user's historical medical records, test results, and diagnostic information by connecting to the hospital information system; the physiological signal acquisition unit collects the user's vital sign data by analyzing individual behavior.

[0005] Specifically, the principal component analysis method performs feature dimensionality reduction and redundancy removal on the multimodal data and integrates the multimodal health data into the multidimensional health set.

[0006] Specifically, the nearest neighbor feature embedding model generates dynamic risk level intervals by weighted clustering of the temporal features of physiological parameters and the changing trends of medical indicators, and determines the credibility of each risk level by combining the membership function.

[0007] Specifically, the subset retention method divides the collected multimodal health dataset into a training set and a test set according to individuals, and inputs the individual data as training samples into the chronic disease risk classification model to calculate potential chronic disease risks.

[0008] Specifically, the spatial structure described in S2 includes the spatial features and structured graph relationships in multimodal health data, specifically including: S101: The multimodal health data is processed by nonlinear mapping based on a dense recurrent feedforward network to extract the high-dimensional spatial coupling relationship between different health feature variables and form a multimodal spatial feature representation; S102: Construct a structured graph relationship based on chronic disease record information, represent individual health characteristics as nodes in the graph structure, represent the association paths between different diseases in historical medical knowledge as edges, and input a dense recurrent feedforward network to extract the structural context embedding; S103: The multimodal spatial features and graph structure embedding features are fused together and used as input data for the chronic disease risk classification model to improve the expressive ability and prediction accuracy of the chronic disease risk classification model for chronic disease risk features.

[0009] Specifically, the chronic disease risk classification model, based on a feature importance analysis mechanism, identifies multidimensional health characteristics related to the impact of chronic disease risk, including: S201: Introduce a feature importance analysis mechanism to assess the correlation of input multidimensional health features and screen feature variables that are highly correlated with chronic disease risk; S202: The aforementioned feature variables are combined with principal component analysis dimensionality reduction methods to establish a multimodal spatial structure and chronic disease knowledge graph; S203: Based on the aforementioned multimodal spatial structure and chronic disease knowledge graph, an intervention strategy is generated, providing accurate decision-making basis for triggering risk warning signals. The specific calculation formula is as follows: , , Among them, I(f) i） For the i-th feature f i Importance score, N is the total number of training samples, L(y j Let y be the loss function of the model for the j-th sample, where y j For real labels, f ij Let wi be the i-th feature value of the j-th sample, and f be the feature value. i Normalized weights, Var(f) i ) is a feature f i The variance is M, where M is the total number of features.

[0010] Specifically, the multi-mode risk convolution mechanism analyzes the temporal fluctuation probability distribution of the risk level, calculates the posterior probability of the current risk state, and dynamically adjusts the early warning triggering conditions according to the risk level interval.

[0011] Specifically, the risk warning signal is sent through the health data collection and interaction module, which is integrated into the medical platform and is automatically triggered based on the detected risk status.

[0012] Specifically, the multidimensional health set constructs static indicators and dynamic trends of individual health status by analyzing key feature variables from different sources in multimodal health data.

[0013] The beneficial effects of this invention are as follows: This invention provides an automatic early warning method for chronic disease risk based on multimodal data analysis. By integrating wearable devices, terminal devices, and medical platforms into a multimodal data acquisition system, it achieves comprehensive collection of multi-source heterogeneous health information such as user physiological signals, behavioral data, medical images, and historical medical records. Through feature dimensionality reduction techniques such as principal component analysis, a multidimensional health set is extracted and constructed, providing a rich foundation of individual health status for subsequent risk analysis. The combined model of convolutional neural networks and graph neural networks can capture spatial local features and cross-modal structural features in multimodal data, respectively. At the same time, the user's historical chronic disease records are introduced as a modeling reference to achieve accurate modeling and classification of chronic disease risks in an individualized context. Compared with traditional methods based on a single data source or shallow models, this approach can significantly improve the model's ability to identify complex health states. The leave-one-out validation strategy is suitable for data scenarios with small sample sizes and multivariate features, effectively improving the generalization ability and stability of the risk prediction model. By combining fuzzy clustering analysis and variational Bayesian inference mechanisms, this invention can not only identify static chronic disease risks, but also dynamically perceive the fluctuation trend of risk levels and has adaptive early warning capabilities. Finally, this invention realizes real-time monitoring and intelligent early warning of chronic disease risks, supports personalized intervention suggestion push, and can be linked with doctors' terminals, thus possessing good clinical application value and promotion prospects. Attached Figure Description

[0014] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating an automatic early warning method for chronic disease risk based on multimodal data analysis according to the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0017] Please see Figure 1 An automatic early warning method for chronic disease risk based on multimodal data analysis: S1: Collect users' multimodal health data in real time through the multimodal data acquisition module, and perform dimensionality reduction processing on the multimodal health data based on the principal component analysis method to construct an individual's multidimensional health set; S2: Based on the extracted multidimensional health set, the multimodal data is input into a recurrent feedforward network to extract spatial features. At the same time, the chronic disease record information is transformed into a graph structure of nodes and edges to train the chronic disease risk classification model. The learning rate of the parameters is adjusted by an adaptive matrix estimator. The collected multimodal health data is input into the chronic disease risk classification model. The average classification accuracy and absolute error are calculated based on leave-one-out validation to analyze potential chronic disease risks. S3: Based on the potential chronic disease risk, a risk level range is constructed by combining the individual's current physiological parameters and medical indicator change trends through a nearest neighbor feature embedding model, and multimodal risk intervention information is generated based on the risk level range. S4: The multimodal risk intervention information uses a multimodal risk convolution mechanism, combined with the risk level fluctuation trend, to determine the current risk status using multidimensional health features and the multimodal risk intervention information as an observation sequence. If the current risk status is detected to exceed the safe risk level range, a risk warning signal is automatically issued through the medical platform.

[0018] Specifically, the multimodal data acquisition module includes: a medical record access unit and a physiological signal acquisition unit. The medical record access unit accesses the user's historical medical records, test results, and diagnostic information by connecting to the hospital information system; the physiological signal acquisition unit collects the user's vital sign data by analyzing individual behavior.

[0019] Specifically, the principal component analysis method performs feature dimensionality reduction and redundancy removal on the multimodal data and integrates the multimodal health data into the multidimensional health set.

[0020] Specifically, the nearest neighbor feature embedding model generates dynamic risk level intervals by weighted clustering of the temporal features of physiological parameters and the changing trends of medical indicators, and determines the credibility of each risk level by combining the membership function.

[0021] In the home scenario implementation, during the initial deployment, users need to equip their homes with a smart health box that includes multimodal data acquisition capabilities, such as an electronic blood pressure monitor, blood glucose meter, pulse oximeter with Wi-Fi communication capabilities, and a health record app connected to their smartphones. This smart system collects the user's basic physiological parameters, such as blood pressure, fasting blood glucose, and heart rate, at set times every morning and evening. Combined with manually entered diet, sleep, and mood scores from the app, it constructs a raw health feature dataset. An automatic modeling module normalizes and removes noise from the continuous data uploaded by the user, and then uses principal component analysis (PCA) to compress and extract data dimensions, retaining features highly correlated with diseases, resulting in an individual multidimensional health set. This set semantically describes the user's physiological stability, behavioral regularity, and psychological stress intensity, and utilizes a CNN-GNN joint modeling architecture for chronic disease risk classification. The CNN sub-network extracts data from the collected time-series data using sliding window convolution to capture short-term trends in blood pressure changes; the GNN sub-network constructs a feature graph, connecting feature nodes according to relevance edge weights to model the graph-structured dependency relationship between blood glucose and sleep quality. The model output is used for chronic disease risk grading. A leave-one-out validation strategy is employed to calibrate the model's generalization ability. Fuzzy clustering analysis is performed based on daily risk prediction results to determine the user's risk level range, categorizing it into three types: "healthy homeostasis," "alert transitional state," and "borderline high-risk state." The platform then generates personalized intervention suggestions based on the user's risk level. For example, if a user's systolic blood pressure is detected to be elevated for three consecutive days, accompanied by decreased sleep efficiency, the system will push lifestyle suggestions, recommend nutritional consultation services, or suggest remote doctor appointments for online initial screening. To prevent misjudgments due to individual state fluctuations, the system introduces a variational Bayesian inference mechanism, combining historical health status sequences with the model's prediction uncertainty distribution to dynamically assess the current prediction confidence. When a certain characteristic signal shows a non-periodic and drastic deviation, the system will identify it as a "mutation risk event" and immediately push a red alert through the app, while also allowing users to manually provide feedback on the actual situation for further optimization of model parameters.

[0022] Specifically, the subset retention method divides the collected multimodal health dataset into a training set and a test set according to individuals, and inputs the individual data as training samples into the chronic disease risk classification model to calculate potential chronic disease risks.

[0023] Specifically, the spatial structure described in S2 includes the spatial features and structured graph relationships in multimodal health data, specifically including: S101: The multimodal health data is processed by nonlinear mapping based on a dense recurrent feedforward network to extract the high-dimensional spatial coupling relationship between different health feature variables and form a multimodal spatial feature representation; S102: Construct a structured graph relationship based on chronic disease record information, represent individual health characteristics as nodes in the graph structure, represent the association paths between different diseases in historical medical knowledge as edges, and input a dense recurrent feedforward network to extract the structural context embedding; S103: The multimodal spatial features and graph structure embedding features are fused together and used as input data for the chronic disease risk classification model to improve the expressive ability and prediction accuracy of the chronic disease risk classification model for chronic disease risk features.

[0024] Specifically, the chronic disease risk classification model, based on a feature importance analysis mechanism, identifies multidimensional health characteristics related to the impact of chronic disease risk, including: S201: Introduce a feature importance analysis mechanism to assess the correlation of input multidimensional health features and screen feature variables that are highly correlated with chronic disease risk; S202: The aforementioned feature variables are combined with principal component analysis dimensionality reduction methods to establish a multimodal spatial structure and chronic disease knowledge graph; S203: Based on the aforementioned multimodal spatial structure and chronic disease knowledge graph, an intervention strategy is generated, providing accurate decision-making basis for triggering risk warning signals. The specific calculation formula is as follows: , , Among them, I(f) i） For the i-th feature f i Importance score, N is the total number of training samples, L(y j Let y be the loss function of the model for the j-th sample, where y j For real labels, f ij Let wi be the i-th feature value of the j-th sample, and f be the feature value. i Normalized weights, Var(f) i ) is a feature f i The variance is M, where M is the total number of features.

[0025] Specifically, the multi-mode risk convolution mechanism analyzes the temporal fluctuation probability distribution of the risk level, calculates the posterior probability of the current risk state, and dynamically adjusts the early warning triggering conditions according to the risk level interval.

[0026] In the community health service center implementation, a multimodal data acquisition module is deployed. The community health service center provides elderly users participating in the project with wearable health monitoring bracelets and installs a data synchronization app on their smartphones. The user's real-time physiological data is uploaded to the mobile device via Bluetooth and periodically synchronized to a cloud-based health platform. Simultaneously, the app guides users to record their daily sleep time, meal times, and medication records. The system automatically imports their past medical history, test reports, and drug allergy history from the national regional health record platform, forming a preliminary individual multimodal health database. Principal component analysis is used to reduce the dimensionality of the collected multimodal health data, removing information redundancy, highlighting potential influencing factors, and extracting a multidimensional health set for the user, covering parameters such as cardiovascular indicators, behavioral patterns, and time stability. This multidimensional health set is then input into a chronic disease risk classification model constructed by fusing convolutional neural networks (CNNs) and graph neural networks (GNNs). The CNN is responsible for modeling the spatial trends of time-series physiological characteristics, while the GNN identifies potential structural relationships between blood pressure fluctuations, glucose tolerance, and historical disease diagnoses by constructing a user health feature map. During the training phase, the model integrates historical data and uses leave-one-out cross-validation to assess generalization ability. It outputs the user's current potential chronic disease risk level and uses a fuzzy C-means algorithm to perform cluster analysis on the risk results, classifying users into low-risk, medium-risk, or high-risk categories. The system then generates personalized early warning reports and dynamically adjusts intervention strategies based on risk level and trend changes. For example, if a user's blood pressure does not decrease at night, their heart rate increases, and they have a history of diabetes, the system will automatically identify them as a "potentially high-risk" state and push an early warning report to the community doctor's platform via the medical cloud platform, prompting them to arrange a follow-up examination or adjust their medication plan.

[0027] Specifically, the multidimensional health set constructs static indicators and dynamic trends of individual health status by analyzing key feature variables from different sources in multimodal health data.

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An automatic early warning method for chronic disease risk based on multimodal data analysis, characterized in that, include: S1: Collect users' multimodal health data in real time through the multimodal data acquisition module, and perform dimensionality reduction processing on the multimodal health data based on the principal component analysis method to construct an individual's multidimensional health set; S2: Based on the extracted multidimensional health set, the multimodal data is input into a dense recurrent feedforward network to extract spatial features. At the same time, the chronic disease record information is transformed into a graph structure of nodes and edges to train the chronic disease risk classification model. The learning rate of the parameters is adjusted by an adaptive matrix estimator. The collected multimodal health data is input into the chronic disease risk classification model. The average classification accuracy and absolute error are calculated based on the subset retention method to analyze potential chronic disease risks. S3: Based on the potential chronic disease risk, a risk level range is constructed by combining the individual's current physiological parameters and medical indicator change trends through a nearest neighbor feature embedding model, and multimodal risk intervention information is generated based on the risk level range. S4: The multimodal risk intervention information uses a multimodal risk convolution mechanism, combined with the risk level fluctuation trend, to determine the current risk status using multidimensional health features and the multimodal risk intervention information as an observation sequence. If the current risk status is detected to exceed the safe risk level range, a risk warning signal is automatically issued through the medical platform.

2. The system according to claim 1, characterized in that, The multimodal data acquisition module includes: a medical record access unit and a physiological signal acquisition unit. The medical record access unit accesses the user's historical medical records, test results, and diagnostic information by connecting to the hospital information system. The physiological signal acquisition unit collects the user's vital sign data by analyzing individual behavior.

3. The system according to claim 1, characterized in that, The principal component analysis method performs feature dimensionality reduction and redundancy removal on multimodal data and integrates the multimodal health data into the multidimensional health set.

4. The system according to claim 1, characterized in that, The spatial structure described in S2 includes the spatial features and structured graph relationships in multimodal health data, specifically including: S101: The multimodal health data is processed by nonlinear mapping based on a dense recurrent feedforward network to extract the high-dimensional spatial coupling relationship between different health feature variables and form a multimodal spatial feature representation; S102: Construct a structured graph relationship based on chronic disease record information, represent individual health characteristics as nodes in the graph structure, represent the association paths between different diseases in historical medical knowledge as edges, and input a dense recurrent feedforward network to extract the structural context embedding; S103: The multimodal spatial features and graph structure embedding features are fused together and used as input data for the chronic disease risk classification model to improve the expressive ability and prediction accuracy of the chronic disease risk classification model for chronic disease risk features.

5. The system according to claim 1, characterized in that, The subset retention method divides the collected multimodal health dataset into training and testing sets according to individuals, and uses the individual data as training samples to input into the chronic disease risk classification model to calculate potential chronic disease risks.

6. The system according to claim 3, characterized in that, The nearest neighbor feature embedding model generates dynamic risk level intervals by weighted clustering of the temporal features of physiological parameters and the changing trends of medical indicators, and determines the credibility of each risk level by combining the membership function.

7. The method according to claim 4, characterized in that, The chronic disease risk classification model identifies multidimensional health characteristics related to the impact of chronic disease risk based on a feature importance analysis mechanism, including: S201: Introduce a feature importance analysis mechanism to assess the correlation of input multidimensional health features and screen feature variables that are highly correlated with chronic disease risk; S202: The aforementioned feature variables are combined with principal component analysis dimensionality reduction methods to establish a multimodal spatial structure and chronic disease knowledge graph; S203: Generate intervention strategies based on the aforementioned multimodal spatial structure and chronic disease knowledge graph, and provide accurate decision-making basis for triggering risk warning signals.

8. The system according to claim 6, characterized in that, The multi-mode risk convolution mechanism analyzes the temporal fluctuation probability distribution of the risk level, calculates the posterior probability of the current risk state, and dynamically adjusts the early warning triggering conditions according to the risk level interval.

9. The system according to claim 1, characterized in that, The risk warning signal is sent through the health data acquisition and interaction module, which is integrated into the medical platform and is automatically triggered based on the detected risk status.

10. The system according to claim 4, characterized in that, The multidimensional health dataset constructs static indicators and dynamic trends of an individual's health status by analyzing key feature variables from different sources in multimodal health data.

Images