Health management data mining method and system based on deep learning

Abstract

The invention provides a health management data mining method and system based on deep learning, and the method comprises the steps: firstly obtaining a target health data set, including physiological monitoring, behavior activity records and medical history texts, of a target user, and then carrying out the cross-modal noise filtering and feature alignment processing of the target health data set; the method comprises the following steps: generating a standardized health feature sequence containing a synchronous timestamp and the like, performing hierarchical feature abstraction by using parallel time sequence convolution branches and space attention branches through a cascaded spatial-temporal feature extraction network to obtain a multi-granularity health state characterization vector, inputting the multi-granularity health state characterization vector into a pre-training model, and training the multi-granularity health state characterization vector to obtain a multi-granularity health state characterization model. According to the method, a personalized health management scheme containing disease risk levels and the like is generated, finally, according to dynamic monitoring period parameters in the personalized health management scheme, the data acquisition frequency and the feature alignment strategy are adjusted in real time, a closed-loop health management data flow is formed, and precise and personalized health management is achieved.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and more specifically, to a health management data mining method and system based on deep learning. Background Art

[0002] As people pay more and more attention to health, the field of health management has developed rapidly. However, there are many limitations in the existing health management technology, which makes it difficult for the effect and accuracy of health management to meet the growing demand.

[0003] Early health management mainly relied on manual records and simple statistical analysis of data. Medical staff manually collected basic health information of patients, such as height, weight, blood pressure and other single-dimensional data, and then made a preliminary assessment of the patient's health status through experience and simple calculation formulas. This method is not only inefficient, but also difficult to accurately grasp the patient's overall health status and detect potential health risks in advance due to the incompleteness of data collection and the simplicity of analysis methods.

[0004] With the development of information technology, the existing information technology has limited processing capabilities for multi-source heterogeneous data, which makes it impossible to fully explore the potential value of the data, limiting the accuracy and effectiveness of health management decisions.

[0005] In terms of health risk prediction, traditional methods are mostly based on fixed models and preset rules, lacking consideration of the personalized characteristics of different users. Using the same risk prediction model and management plan for users with different living habits, genetic backgrounds, and health conditions leads to inaccurate prediction results, lack of targeted management plans, and failure to truly meet users' personalized health management needs.

[0006] In addition, the existing health management system lacks a dynamic adjustment mechanism. Once the data collection frequency and management strategy are set, it is difficult to flexibly adjust according to the real-time changes in the user's health status. For example, for users with stable health conditions, data is still collected at a higher frequency, which not only wastes resources but also increases the burden on users; for users whose health conditions fluctuate, if the data collection frequency and management strategy cannot be adjusted in time, the best time for intervention may be missed, affecting the effectiveness of health management. Summary of the invention

[0007] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a health management data mining method based on deep learning, the method comprising: Acquire a target health data set of a target user, wherein the target health data set includes physiological monitoring data, behavioral activity record data, and medical history text data, wherein the physiological monitoring data includes multiple types of continuous physiological parameters, the behavioral activity record data includes movement status and sleep quality indicators in a time series, and the medical history text data includes structured diagnosis records and unstructured symptom descriptions; Performing cross-modal noise filtering and feature alignment processing on the target health data set to generate a standardized health feature sequence, wherein each feature unit in the standardized health feature sequence includes a synchronization timestamp, a cross-modal association identifier, and a normalized numerical expression; The standardized health feature sequence is hierarchically abstracted through a cascaded spatiotemporal feature extraction network to obtain a multi-granularity health state representation vector, wherein the spatiotemporal feature extraction network includes a temporal convolution branch and a spatial attention branch arranged in parallel, wherein the temporal convolution branch is used to capture long-term health trends, and the spatial attention branch is used to identify dynamic dependencies between cross-modal features; Inputting the multi-granularity health status representation vector into a pre-trained health risk prediction model to generate a personalized health management plan for the target user, wherein the personalized health management plan includes a disease risk level label, an intervention measure priority list, and a dynamic monitoring cycle configuration parameter; According to the dynamic monitoring cycle configuration parameters in the personalized health management solution, the collection frequency and feature alignment strategy of the target health data set are adjusted in real time to form a closed-loop health management data flow.

[0008] On the other hand, an embodiment of the present application also provides a health management platform system, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0009] Based on the above aspects, the embodiment of the present application comprehensively collects a multi-source heterogeneous target health data set of the target user including physiological monitoring data, behavioral activity record data and medical history text data, and performs cross-modal noise filtering and feature alignment processing to generate a standardized health feature sequence. Each feature unit contains a synchronization timestamp, a cross-modal association identifier and a normalized numerical expression. The data of different modalities are closely associated through the cross-modal association identifier, and the synchronization timestamp further enhances the consistency of the data in the time dimension. Furthermore, the parallel temporal convolution branch and the spatial attention branch work together to capture long-term health trends and identify dynamic dependencies between cross-modal features, respectively. It can deeply mine data features in different dimensions, grasp the long-term changes of health data from the time dimension, and analyze the complex interaction relationship between different modal data from the spatial dimension, thereby obtaining a multi-granularity health status representation vector. Then, a personalized health management plan is generated based on the multi-granularity health status representation vector, which includes disease risk level labels, intervention measure priority lists, and dynamic monitoring cycle configuration parameters. The disease risk level labels help users to clearly understand their own health risk status, and the intervention measure priority list provides users and medical staff with clear health management priorities and directions. The monitoring cycle can be dynamically adjusted through the dynamic monitoring cycle configuration parameters, which can flexibly and intelligently adjust the data collection frequency and feature alignment strategy according to the user's real-time health status changes, forming a closed-loop health management data flow, while ensuring the health management effect, effectively reducing data collection costs and resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a schematic diagram of the execution flow of the health management data mining method based on deep learning provided in an embodiment of the present application.

[0011] Figure 2 It is a schematic diagram of the hardware architecture of the health management platform system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0012] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a health management data mining method based on deep learning provided by an embodiment of the present application. The health management data mining method based on deep learning is introduced in detail below.

[0013] Step S110, obtaining a target health data set of the target user, wherein the target health data set includes physiological monitoring data, behavioral activity record data, and medical history text data, wherein the physiological monitoring data includes multiple types of continuous physiological parameters, the behavioral activity record data includes movement status and sleep quality indicators in a time series, and the medical history text data includes structured diagnosis records and unstructured symptom descriptions.

[0014] In this embodiment, taking the target user as a middle-aged male as an example, his target health data set can be obtained from multiple channels.

[0015] In detail, for physiological monitoring data, various types of continuous physiological parameters can be continuously collected through wearable devices (such as smart bracelets and smart body fat scales). For example, a smart bracelet can monitor heart rate in real time. During the day, his heart rate may be relatively stable when he wakes up in the morning, about 60-70 beats / minute; when exercising (such as brisk walking or climbing stairs), the heart rate will rise to 100-120 beats / minute; and it will fall back to a lower level when resting at night. Smart body fat scales can measure parameters such as weight, body fat percentage, and muscle mass. His weight may be around 75 kg and his body fat percentage is 25%. The above continuous physiological parameters constitute part of the physiological monitoring data.

[0016] In terms of behavioral activity record data, the movement status is recorded through the motion sensor in the mobile phone. Assuming that he has a period of brisk walking every morning, the motion sensor can record his movement status information such as number of steps, step frequency, and movement speed. During sleep, the smart mattress or smart bracelet can monitor sleep quality indicators such as sleep duration, deep sleep duration, light sleep duration, and the number of tossing and turning during sleep. For example, his sleep duration every night is about 7-8 hours, of which the deep sleep duration is about 2-3 hours.

[0017] Medical history text data contains structured diagnosis records and unstructured symptom descriptions. For example, he had a physical examination and disease diagnosis in the hospital before. The structured diagnosis record may show that he was diagnosed with mild hypertension with a blood pressure of about 140 / 90 mmHg. Unstructured symptom descriptions may include occasional dizziness, especially when standing for a long time or standing up suddenly. The above information is recorded in the medical history text data. Therefore, by acquiring data from these different channels, a complete target health data set for the target user is generated.

[0018] Step S120, performing cross-modal noise filtering and feature alignment processing on the target health data set to generate a standardized health feature sequence, wherein each feature unit in the standardized health feature sequence includes a synchronization timestamp, a cross-modal association identifier, and a normalized numerical expression.

[0019] Continuing with the example of the middle-aged man mentioned above, when processing the target health data set, the first step is to detect abnormal fluctuations in the continuous physiological parameters in the physiological monitoring data. For example, the heart rate data may occasionally fluctuate abnormally because the smart bracelet may be affected by external electromagnetic interference or loose wearing during the signal acquisition process. Therefore, an adaptive threshold segmentation algorithm can be used to identify these signal acquisition noise segments. Assuming that the normal heart rate fluctuation range is set to fluctuate up and down by 10 times / minute in a calm state, when the heart rate suddenly jumps to 150 times / minute at a certain moment and lasts for a very short time, the algorithm identifies it as a noise segment. Then, the signal acquisition noise segment is repaired through bidirectional cyclic interpolation, so that the repaired heart rate data is smoother and more accurate.

[0020] The motion type of the motion state indicators in the behavioral activity record data is classified, and his brisk walking exercise every morning is classified as an aerobic fitness action. Then, based on the preset motion energy consumption mapping table, the discrete action event of brisk walking can be converted into a continuous energy consumption curve. Assuming that the energy consumption of brisk walking is related to the step frequency and speed, the energy consumption value of each minute is obtained through the existing calculation formula to form a continuous energy consumption curve. This energy consumption curve is then aligned with the sleep quality indicator in the time window, for example, with 24 hours as a time window, so that the motion state and sleep quality are considered in the same time frame to obtain the aligned behavioral activity curve.

[0021] When processing unstructured symptom descriptions in medical history text data, entity relationship extraction is performed. For his occasional dizziness symptom, the relationship between dizziness and mild hypertension is extracted to construct a triple knowledge graph of symptoms-diagnosis-treatment. For example, a triple relationship such as dizziness (symptom)-mild hypertension (diagnosis)-drug therapy (treatment). Then, the graph nodes in the triple knowledge graph are embedded as low-dimensional semantic vectors, and concepts such as dizziness and mild hypertension are converted into semantic vectorized medical knowledge that can be processed by computers.

[0022] Finally, the repaired physiological parameters, aligned behavioral activity curves, and semantically vectorized medical knowledge are subjected to multimodal feature fusion. Since data of different modalities may be offset in time, for example, physiological monitoring data is collected continuously in real time, while medical history text data is a record of a certain period of time in the past, a dynamic time warping algorithm is used to eliminate the time offset of the cross-modal data and generate a standardized health feature sequence with a unified time reference. Each feature unit contains a synchronized timestamp (such as 9:00 a.m. on August 1, 2024), a cross-modal association identifier (used to identify which modal data the feature comes from, such as physiological monitoring, behavioral activity, or medical history) and a normalized numerical expression (the value after normalizing data of different ranges, such as heart rate normalized to between 0-1).

[0023] Step S130, hierarchical feature abstraction is performed on the standardized health feature sequence through a cascaded spatiotemporal feature extraction network to obtain a multi-granularity health status representation vector. The spatiotemporal feature extraction network includes a temporal convolution branch and a spatial attention branch arranged in parallel. The temporal convolution branch is used to capture long-term health trends, and the spatial attention branch is used to identify dynamic dependencies between cross-modal features.

[0024] Still taking the standardized health feature sequence of this middle-aged man as an example, it is input into the spatiotemporal feature extraction network.

[0025] In the temporal convolution branch, a stacked structure of dilated causal convolutional layers is used to extract multi-scale temporal patterns from standardized health feature sequences. Convolutional layers with different dilation rates play different roles. For example, convolutional layers with a smaller dilation rate can capture short-term physiological fluctuations, such as the rapid increase and then rapid decrease of his heart rate after a short period of exercise. Such short-term fluctuations can be detected; convolutional layers with a moderate dilation rate are used to capture medium-term behavioral patterns, such as the changing trends of his exercise time and sleep quality in the past week; convolutional layers with a larger dilation rate can capture long-term health trends, such as the overall changing trends of his weight and blood pressure in the past few months, thereby obtaining a multi-scale temporal feature map.

[0026] In the spatial attention branch, a cross-modal feature association matrix is ​​constructed. Taking the heart rate and blood pressure in his physiological monitoring data, the exercise energy consumption in the behavioral activity recording data, and the diagnosis of mild hypertension in the medical history text data as examples, the contribution of these different modal features to the health status is calculated through a learnable attention weight allocation mechanism. For example, the heart rate and blood pressure may have a higher weight on his health status because he suffers from mild hypertension, while the exercise energy consumption is also related to controlling blood pressure and improving overall health status. The inter-modal dependency graph is generated through calculation.

[0027] The multi-scale temporal feature map output by the temporal convolution branch is concatenated with the inter-modal dependency graph generated by the spatial attention branch. Assuming that the multi-scale temporal feature map contains feature information such as heart rate, blood pressure, exercise and sleep at different time scales, and the inter-modal dependency graph contains the association weight information between the features of each modality, the two are concatenated and the feature weights are adjusted through the gated fusion unit. For example, if the impact of blood pressure on health status is more critical at the current stage, the weight of the features related to blood pressure will be appropriately increased during fusion. Then, the adjusted features are input into the bidirectional gated recurrent network for context-aware feature enhancement, so that the network can better use the context information to understand the health status features. After obtaining the enhanced fusion features, a hierarchical pooling operation is performed on them. First, a global health status summary vector is extracted, which can reflect his overall health status, such as the overall health risk level is medium; at the same time, a local abnormal pattern feature vector is extracted, such as the local abnormality of occasional dizziness that can be highlighted. Finally, a dimensionality reduction integration is performed through the fully connected layer to generate a multi-granular health status representation vector containing multi-level health information.

[0028] Step S140, input the multi-granularity health status representation vector into a pre-trained health risk prediction model to generate a personalized health management plan for the target user, wherein the personalized health management plan includes a disease risk level label, an intervention measure priority list, and a dynamic monitoring cycle configuration parameter.

[0029] In detail, the multi-granularity health status representation vector of the above middle-aged male can be input into the pre-trained health risk prediction model.

[0030] In the multi-head risk prediction module, the cardiovascular disease risk probability, metabolic syndrome risk index and mental stress assessment score are calculated in parallel. For example, because he suffers from mild hypertension, his cardiovascular disease risk probability is relatively high, assuming that the calculated value is 30%. For the metabolic syndrome risk index, considering factors such as his weight and body fat percentage, the calculated risk index may be at a medium level, such as 40%. In terms of mental stress assessment, by analyzing his behavioral activity records (such as whether he has regular exercise and adequate sleep) and symptom descriptions in his medical history (dizziness may be related to mental stress), the assessment score may be 20%. Each risk prediction module uses an independent feature transformation layer and risk classifier, but shares the underlying features of the multi-granularity health state representation vector.

[0031] The severity of the disease risk level label is determined based on the weighted combination of the cardiovascular disease risk probability and the metabolic syndrome risk index. Assuming that the weight of the cardiovascular disease risk probability is 0.6, the weight of the metabolic syndrome risk index is 0.4, and the weighted combination result is 34%, after dynamically calibrating the grading result in combination with the mental stress assessment score, his disease risk level is determined to be medium risk.

[0032] In the intervention strategy generation module, matching candidate intervention items are retrieved from the preset intervention measures knowledge base based on the calibrated disease risk level. Since he is at medium risk and suffers from mild hypertension, candidate intervention items may include dietary adjustments (such as reducing salt intake), exercise prescriptions (increasing aerobic exercise time and intensity), and medical examination plans (regularly measuring blood pressure, conducting blood lipid tests, etc.). Then the execution feasibility score of each intervention item is calculated based on his historical behavioral activity record data. For example, if he has a certain exercise habit before, the execution feasibility score of the exercise prescription may be high. The reinforcement learning strategy is used to prioritize the candidate intervention items, and a priority list of intervention measures is generated, including dietary adjustment recommendations (such as no more than 6 grams of salt intake per day), exercise prescriptions (increasing 30 minutes of brisk walking exercise time per week), and medical examination plans (a comprehensive physical examination every three months). At the same time, the dynamic monitoring cycle configuration parameters are derived based on the risk probability change rate. Since his risk is at a medium level and relatively stable, the dynamic monitoring cycle is configured to collect his physiological monitoring data and behavioral activity record data once every two weeks, and update and review his medical history data every three months.

[0033] Step S150, according to the dynamic monitoring cycle configuration parameters in the personalized health management plan, the collection frequency and feature alignment strategy of the target health data set are adjusted in real time to form a closed-loop health management data flow.

[0034] In detail, according to the personalized health management plan developed for middle-aged men, when data is collected according to the dynamic monitoring cycle, adjustments are made according to the severity classification of the disease risk level label.

[0035] When his disease risk level is medium risk, physiological monitoring data and behavioral activity record data are collected every two weeks. For example, smart bracelets and smart body fat scales collect data at set time intervals, and motion sensors continuously record his motion status during his daily activities. If at a certain stage, the severity level of his disease risk level label is monitored to exceed the preset threshold (assuming that the preset threshold is 50%), for example, due to a sudden increase in his blood pressure or the emergence of new symptoms such as palpitations, the sampling interval of the physiological monitoring data is shortened to a preset ratio of the original sampling interval (such as shortened to half of the original). The comprehensive physiological data that was originally collected every two weeks is now collected once a week. At the same time, according to the set frequency increase strategy, the number of activations of the motion posture sensor in the behavioral activity record data is increased, for example, from recording the motion status every day to recording once every half a day, so as to obtain his exercise and health status changes in a more timely manner.

[0036] The standardized health feature sequence is dynamically updated based on the multimodal data stream collected in real time. When new data arrives, the sliding window mechanism is used to perform incremental feature alignment on the newly added data segments. For example, the newly collected multimodal data of the previous week is feature aligned with the previous historical data, and the new multimodal data is integrated into the standardized health feature sequence while maintaining the contextual relevance of the historical feature sequence.

[0037] The updated standardized health feature sequence is input into the spatiotemporal feature extraction network for online feature abstraction, and the real-time reasoning process of the health risk prediction model is triggered to generate an updated personalized health management plan. If after a period of adjustment, his health condition improves and the disease risk level decreases, when the disease risk level label of a preset number of monitoring cycles (assuming three cycles) maintains a stable state, the data collection interval is gradually extended until the baseline frequency is restored. For example, physiological data is gradually restored from being collected once a week to once every two weeks, while reducing the intensity of medical examination plans in the priority list of intervention measures, such as changing from a comprehensive physical examination every three months to once every six months. In this way, a closed-loop health management data stream is formed, which adjusts the data collection and management strategies in real time according to the user's health status, and continuously optimizes personalized health management plans.

[0038] Based on the above steps, the embodiment of the present application comprehensively collects a multi-source heterogeneous target health data set of the target user including physiological monitoring data, behavioral activity record data and medical history text data, and performs cross-modal noise filtering and feature alignment processing to generate a standardized health feature sequence. Each feature unit contains a synchronization timestamp, a cross-modal association identifier and a normalized numerical expression. The data of different modalities are closely associated through the cross-modal association identifier, and the synchronization timestamp further enhances the consistency of the data in the time dimension. Furthermore, the parallel temporal convolution branch and the spatial attention branch work together to capture long-term health trends and identify dynamic dependencies between cross-modal features, respectively. It can deeply mine data features in different dimensions, grasp the long-term change law of health data from the time dimension, and analyze the interaction relationship between different modal data from the spatial dimension, thereby obtaining a multi-granularity health status representation vector. Then, a personalized health management plan is generated based on the multi-granularity health status representation vector, which includes disease risk level labels, intervention measure priority lists, and dynamic monitoring cycle configuration parameters. The disease risk level labels help users to clearly understand their own health risk status, and the intervention measure priority list provides users and medical staff with clear health management priorities and directions. The monitoring cycle can be dynamically adjusted through the dynamic monitoring cycle configuration parameters, which can flexibly and intelligently adjust the data collection frequency and feature alignment strategy according to the user's real-time health status changes, forming a closed-loop health management data flow, while ensuring the health management effect, effectively reducing data collection costs and resource consumption.

[0039] In a possible implementation, step S120 specifically includes: Step S121, abnormal fluctuation detection is performed on the continuous physiological parameters in the physiological monitoring data, an adaptive threshold segmentation algorithm is used to identify the signal acquisition noise segment, and data repair is performed on the signal acquisition noise segment through bidirectional cyclic interpolation to obtain the repaired physiological parameters.

[0040] In detail, for the continuous physiological parameters in the physiological monitoring data of the middle-aged male in the above example, such as heart rate data, there may be abnormal fluctuations due to the influence of equipment or external factors. Under normal circumstances, the heart rate of a middle-aged male in a quiet state is in the range of 60-70 times / minute, which may rise to 80-90 times / minute during daily activities such as walking, and may reach 100-120 times / minute when the exercise is more intense. Then, an adaptive threshold segmentation algorithm is used to identify the signal acquisition noise segment, and the heart rate fluctuation threshold in the quiet state is set to 5 times / minute. When the heart rate data collected at a certain moment is 150 times / minute and the surrounding data are all in the normal fluctuation range, the adaptive threshold segmentation algorithm determines that the data segment is a signal acquisition noise segment. After that, the signal acquisition noise segment is repaired by bidirectional cyclic interpolation. For example, based on the normal heart rate data before and after the noise segment, it is repaired according to a certain interpolation algorithm so that the repaired data and the previous and next data are smoothly transitioned to obtain the repaired physiological parameters.

[0041] Step S122, classify the motion type of the motion state indicator in the behavior activity record data, convert discrete motion events into continuous energy consumption curves based on a preset motion energy consumption mapping table, and align the time window with the sleep quality indicator to obtain an aligned behavior activity curve.

[0042] In terms of behavioral activity record data, for the brisk walking exercise every morning for middle-aged men, this exercise state indicator needs to be classified into action types and classified as aerobic fitness actions. Based on the preset exercise energy consumption mapping table, which contains the relationship between different step frequencies, speeds and other factors and energy consumption, the energy consumption value is calculated according to the step frequency and speed data during his brisk walking. Assuming that his step frequency is 120 steps / minute and the speed is 4 kilometers / hour when he is walking fast, the energy consumption value per minute is calculated by the formula in the mapping table. As time goes by, these discrete energy consumption values ​​are converted into continuous energy consumption curves. At the same time, his sleep quality indicators include information such as sleep duration and deep sleep duration. With 24 hours as a time window, the energy consumption curve is aligned with the sleep quality indicator in the time window. For example, in this 24-hour time window, the energy consumption peak corresponding to the morning brisk walking and indicators such as the sleep duration at night are arranged in time to obtain the aligned behavioral activity curve.

[0043] Step S123, extracting entity relationships from the unstructured symptom descriptions in the medical history text data, constructing a triple knowledge graph of symptom-diagnosis-treatment, and embedding the graph nodes in the triple knowledge graph into low-dimensional semantic vectors to obtain semantically vectorized medical knowledge.

[0044] Entity relationship extraction is performed for unstructured symptom descriptions in the medical history text data of middle-aged men, such as the symptom of occasional dizziness. For example, he was diagnosed with mild hypertension, and a triple knowledge graph of symptoms-diagnosis-treatment was constructed, namely dizziness (symptom)-mild hypertension (diagnosis)-drug therapy (treatment). For the graph nodes in the triple knowledge graph, such as dizziness, mild hypertension, and drug therapy, they are embedded into low-dimensional semantic vectors through a specific embedding algorithm. Taking the concept of dizziness as an example, it is converted into semantically vectorized medical knowledge that can be processed by computers. The low-dimensional semantic vector contains the representation of the semantic information of the symptom of dizziness in a low-dimensional space, allowing the computer to understand and process the relationship between these medical knowledge and other modal data.

[0045] Step S124, multimodal feature fusion is performed on the repaired physiological parameters, aligned behavioral activity curves, and semantically vectorized medical knowledge, and a dynamic time warping algorithm is used to eliminate the time offset of cross-modal data to generate the standardized health feature sequence with a unified time reference.

[0046] In detail, the physiological monitoring data of middle-aged men are collected continuously in real time, and the medical history text data are records of the past, and there are differences in time between the two. Therefore, a dynamic time warping algorithm is needed to eliminate the time offset of this cross-modal data so that all data are under a unified time reference. For example, the diagnosis of mild hypertension in the past medical history is aligned in time with the current heart rate, energy consumption curve and other data to generate a standardized health feature sequence with a unified time reference, where each feature unit contains a synchronized timestamp (such as 9 am on August 1, 2024), a cross-modal association identifier (identifying that the feature comes from physiological monitoring, behavioral activities or medical history) and a normalized numerical expression (the normalized value of data in different ranges, such as heart rate normalized to between 0-1).

[0047] In a possible implementation, step S130 specifically includes: Step S131, in the time series convolution branch, a dilated causal convolution layer stacking structure is used to extract multi-scale time series patterns from the standardized health feature sequence to obtain a multi-scale time series feature graph, wherein convolution layers with different dilation rates capture short-term physiological fluctuations, medium-term behavioral patterns, and long-term health trends, respectively.

[0048] In detail, the standardized health feature sequence of the middle-aged man can be input into the temporal convolution branch of the spatiotemporal feature extraction network, and the multi-scale temporal pattern can be extracted by using the stacked structure of the dilated causal convolution layer. Convolutional layers with different dilation rates play different roles. Convolutional layers with a smaller dilation rate can capture short-term physiological fluctuations. For example, after a middle-aged man climbs stairs for a short period of time, his heart rate will rise rapidly and then drop in a short period of time. This short-term heart rate fluctuation can be accurately detected by the convolutional layer with a smaller dilation rate. Convolutional layers with a moderate dilation rate can capture medium-term behavioral patterns, such as the changing trends of his exercise time and sleep quality in the past week. For example, this week, due to his busy work, his exercise time has decreased by 20% compared with last week, and his sleep quality has also declined. This medium-term behavioral pattern change can be captured by this convolutional layer. Convolutional layers with a larger dilation rate are used to capture long-term health trends, such as the overall trend of his weight and blood pressure in the past few months. His weight has increased by 2 kg in the past three months, and his blood pressure has also tended to rise slightly. In this way, a multi-scale temporal feature map is obtained.

[0049] Step S132, in the spatial attention branch, construct a cross-modal feature association matrix, calculate the contribution of different modal features to the health status through a learnable attention weight allocation mechanism, and generate an inter-modal dependency graph.

[0050] Taking the heart rate and blood pressure in the physiological monitoring data of a middle-aged man, the exercise energy consumption in the behavioral activity record data, and the diagnosis of mild hypertension in the medical history text data as examples, the contribution of these different modal features to the health status is calculated through a learnable attention weight allocation mechanism. Since he suffers from mild hypertension, the weight of heart rate and blood pressure on his health status is relatively high, for example, the weight of heart rate is 0.3, and the weight of blood pressure is 0.4. The exercise energy consumption also has a certain correlation with his overall health status, and the weight may be 0.2. The weight of the diagnosis of mild hypertension in the medical history is 0.1. Through such calculations, an inter-modal dependency graph is generated, which can clearly show the dependency between each modal feature and the importance of its impact on the health status.

[0051] In step S133, the multi-scale temporal feature map output by the temporal convolution branch is feature-concatenated with the inter-modal dependency graph generated by the spatial attention branch. After the feature weights are adjusted by the gated fusion unit, the map is input into the bidirectional gated recurrent network for context-aware feature enhancement to obtain enhanced fused features.

[0052] Assuming that the multi-scale time series feature graph contains feature information such as heart rate, blood pressure, exercise and sleep at different time scales, and the inter-modal dependency graph contains the correlation weight information between the features of each modality, the two are spliced ​​and the feature weights are adjusted through the gated fusion unit. If the impact of blood pressure on health status is more critical at the current stage, the weight of the features related to blood pressure will be appropriately increased during fusion, for example, from the original 0.4 to 0.5. The adjusted features are then input into the bidirectional gated recurrent network for context-aware feature enhancement. The bidirectional gated recurrent network can use the contextual information to better understand the health status features. For example, the health status features are enhanced by combining the previous health status and current status changes of middle-aged men to obtain enhanced fusion features.

[0053] Step S134, performing a hierarchical pooling operation on the enhanced fusion features, extracting the global health status summary vector and the local abnormal pattern feature vector respectively, and performing dimensionality reduction integration through a fully connected layer to generate the multi-granularity health status representation vector containing multi-level health information.

[0054] For example, first extract the global health status summary vector, which can reflect the overall health status of middle-aged men. For example, the overall health risk level is medium, which may be because although he suffers from mild hypertension, the risk has not deteriorated further through certain exercise and diet control. At the same time, extract the local abnormal pattern feature vector, which can highlight the local abnormality of his occasional dizziness. Finally, perform dimensionality reduction integration through the fully connected layer, integrate the information in the global health status summary vector and the local abnormal pattern feature vector, and generate a multi-granular health status representation vector containing multi-level health information. The multi-granular health status representation vector can comprehensively describe the health status of middle-aged men, including both the overall health risk level and the local abnormal feature information.

[0055] In a possible implementation, the pre-trained health risk prediction model includes a multi-head risk prediction module and an intervention strategy generation module, and step S140 specifically includes: Step S141, in the multi-head risk prediction module, the cardiovascular disease risk probability, the metabolic syndrome risk index and the mental stress assessment score are calculated in parallel, wherein each risk prediction module uses an independent feature transformation layer and risk classifier, and shares the underlying features of the multi-granularity health status representation vector.

[0056] In this embodiment, the multi-granularity health status representation vector of middle-aged men is input into the multi-head risk prediction module in the pre-trained health risk prediction model. In the multi-head risk prediction module, the cardiovascular disease risk probability, the metabolic syndrome risk index and the mental stress assessment score are calculated in parallel.

[0057] Regarding the calculation of cardiovascular disease risk probability, since middle-aged men suffer from mild hypertension, their blood pressure is around 140 / 90 mmHg, which is an important risk factor for cardiovascular disease. At the same time, his weight is 75 kg, his body fat percentage is 25%, and he is middle-aged. These factors are combined and calculated through specific algorithms and models. The feature transformation layer in the model transforms the relevant features in the multi-granular health status representation vector, such as normalizing and weighting features such as blood pressure, weight, and age. Then the risk classifier calculates the cardiovascular disease risk probability based on the transformed features. Assume that the calculated result is 30%.

[0058] When calculating the metabolic syndrome risk index, factors such as his weight, body fat percentage, exercise and diet are taken into account. His weight is high and his body fat percentage is also at a high level. Although he has a certain exercise habit, the intensity and duration of exercise may not be enough to completely reduce the risk of metabolic syndrome. These related features are processed through independent feature transformation layers, such as recoding features such as weight, body fat percentage and exercise time, and then calculated by the corresponding risk classifier, and the metabolic syndrome risk index may be 40%.

[0059] The calculation of the mental stress assessment score is based on the relevant information in his behavioral activity record data and medical history text data. From the behavioral activity record data, his sleep quality indicators show that he sleeps for about 2-3 hours a night and sleeps for a long time, which may suggest that he has a certain amount of mental stress. At the same time, occasional dizziness symptoms in medical history may also be related to mental stress. In this multi-head risk prediction module, the feature transformation layer transforms relevant features such as sleep quality and dizziness symptoms, and then the risk classifier calculates that the mental stress assessment score may be 20%. Each risk prediction module uses an independent feature transformation layer and risk classifier, but shares the underlying features of the multi-granularity health status representation vector, which contains basic information about the overall health status of middle-aged men, such as physiological parameters, behavioral activity characteristics, etc.

[0060] Step S142, determining the severity classification of the disease risk level label according to the weighted combination result of the cardiovascular disease risk probability and the metabolic syndrome risk index, and dynamically calibrating the classification result in combination with the mental stress assessment score.

[0061] For example, the weight of the cardiovascular disease risk probability is set to 0.6, and the weight of the metabolic syndrome risk index is set to 0.4. The weighted combination result is (30%×0.6+40%×0.4)=34%. Then the grading results are dynamically calibrated in combination with the mental stress assessment score. Since the mental stress assessment score is 20%, although it is not very high, it will also have a certain impact on the overall health status. After a specific calibration algorithm, the previous 34% result is adjusted, and the disease risk level of middle-aged men is finally determined to be medium risk. This calibration process takes into account the potential impact of mental stress on overall health risks, making the assessment of disease risk levels more comprehensive and accurate.

[0062] Step S143, in the intervention strategy generation module, based on the calibrated disease risk level, matching candidate intervention items are retrieved from the preset intervention measure knowledge base, and the execution feasibility score of each intervention item is calculated according to the historical behavior activity record data of the target user.

[0063] Because middle-aged men have mild hypertension and a moderate disease risk level, candidate interventions include dietary adjustments (such as reducing salt intake), exercise prescriptions (increasing the time and intensity of aerobic exercise), and medical examination programs (regularly measuring blood pressure, performing blood lipid tests, etc.).

[0064] The feasibility score of each intervention item is calculated based on the historical behavioral activity record data of the middle-aged male. He had a certain exercise habit before, such as brisk walking every morning, so the feasibility score of the exercise prescription is relatively high. For dietary adjustment, considering factors such as his eating habits and social activities, if he often eats out, the feasibility score of reducing salt intake may be relatively low. For the medical examination plan, since he had a record of regular blood pressure checks according to the doctor's advice, the feasibility score of this intervention item is high.

[0065] Step S144, using a reinforcement learning strategy to prioritize candidate intervention items, generating a priority list of intervention measures including dietary adjustment suggestions, exercise prescriptions, and medical examination plans, and deriving the dynamic monitoring cycle configuration parameters based on the risk probability change rate.

[0066] In detail, the reinforcement learning strategy will comprehensively consider factors such as the feasibility score of each intervention, the potential impact on reducing disease risk, and long-term health benefits. For example, exercise prescriptions are ranked higher in the priority list because of their high feasibility and positive impact on controlling blood pressure and improving overall health status. Although dietary adjustments are slightly less feasible, they are also very important for controlling hypertension and are ranked after exercise prescriptions. Medical examination plans are ranked third based on their necessity and previous implementation. Finally, a priority list of intervention measures is generated, including dietary adjustment recommendations (such as no more than 6 grams of salt intake per day), exercise prescriptions (increasing brisk walking time by 30 minutes per week), and medical examination plans (comprehensive physical examinations every three months). At the same time, the dynamic monitoring cycle configuration parameters are derived based on the risk probability change rate. Since the risk of middle-aged men is at a medium level and relatively stable, and the risk probability changes slowly, the dynamic monitoring cycle is configured to collect his physiological monitoring data and behavioral activity record data once every two weeks, and update and review his medical history data every three months.

[0067] In a possible implementation, step S150 specifically includes: Step S151, when it is monitored that the severity level of the disease risk level label exceeds a preset threshold, the sampling interval of the physiological monitoring data is shortened to a preset proportion of the original sampling interval, and the number of activations of the motion posture sensor in the behavioral activity record data is increased according to the set frequency increase strategy.

[0068] When the severity level of the disease risk level label of a middle-aged male is monitored to exceed the preset threshold, the preset threshold is assumed to be 50%. If during a certain monitoring period, his blood pressure suddenly rises to 160 / 100 mmHg or new symptoms such as palpitations appear, the disease risk level exceeds 50%. At this time, the sampling interval of physiological monitoring data is shortened to the preset proportion of the original sampling interval. The sampling interval of comprehensive physiological data, which was originally collected every two weeks, is now shortened to half of the original, and is collected once a week. At the same time, according to the set frequency increase strategy, the number of activations of the motion posture sensor in the behavioral activity record data is increased, for example, from recording the motion status every day to recording once every half a day. In this way, his exercise and health status changes can be obtained more timely, so that his health status can be monitored more closely.

[0069] Step S152, dynamically updating the standardized health feature sequence according to the multimodal data stream collected in real time, using a sliding window mechanism to perform incremental feature alignment processing on the newly added data segments, and retaining the contextual relevance of the historical feature sequence.

[0070] When new data arrives, such as a week of newly collected data, a sliding window mechanism is used to perform incremental feature alignment on the newly added data segments. The newly collected data includes new physiological monitoring data (such as heart rate, blood pressure, etc.), behavioral activity record data (such as exercise status, sleep quality, etc.), and possible medical history update data (if there is a new diagnosis or symptom description). Taking physiological monitoring data as an example, the new heart rate data needs to be aligned with the previous heart rate data in time to ensure data consistency. For behavioral activity record data, the new exercise status data must be arranged according to the same time reference as the previous exercise status data. In this way, while maintaining the contextual relevance of the historical feature sequence, the new data is integrated into the standardized health feature sequence, so that the standardized health feature sequence can reflect the latest health status of middle-aged men in real time.

[0071] Step S153, inputting the updated standardized health feature sequence into the spatiotemporal feature extraction network for online feature abstraction, and triggering the real-time reasoning process of the health risk prediction model to generate an updated personalized health management plan.

[0072] The spatiotemporal feature extraction network processes the new standardized health feature sequence and recalculates the multi-granular health state representation vector. For example, the increase in blood pressure in the new physiological monitoring data may be reflected in the multi-granular health state representation vector as a change in the characteristic value related to cardiovascular disease risk. Then the new multi-granular health state representation vector is input into the health risk prediction model again for calculation. The multi-head risk prediction module recalculates the cardiovascular disease risk probability, metabolic syndrome risk index and mental stress assessment score. The cardiovascular disease risk probability may rise to 40% due to the increase in blood pressure. According to the new risk probability and other related indicators, the disease risk level is re-determined, such as becoming a higher risk. The intervention strategy generation module re-retrieves and adjusts the priority list of intervention measures according to the new disease risk level, such as increasing the frequency of medical examinations, adjusting the intensity of exercise prescriptions, etc., and re-derives the dynamic monitoring cycle configuration parameters, such as shortening the comprehensive collection of physiological monitoring data and behavioral activity record data to weekly, and updating and reviewing medical history data every two months.

[0073] Step S154, when the disease risk level label maintains a stable state for a preset number of consecutive monitoring cycles, the data collection interval is gradually extended until the baseline frequency is restored, and the intensity of the medical examination plan in the intervention measure priority list is reduced.

[0074] When the disease risk level label remains stable for a preset number of monitoring cycles (assuming three cycles), for example, after a period of adjustment and treatment, the blood pressure of middle-aged men stabilizes at around 130 / 80 mmHg, and the disease risk level remains at a medium risk level. At this time, the data collection interval is gradually extended until the baseline frequency is restored. Gradually restore from collecting physiological data once a week to collecting it once every two weeks, and at the same time reduce the intensity of the medical examination plan in the priority list of intervention measures, such as changing from a comprehensive physical examination every two months to once every six months. Therefore, under the premise of ensuring effective monitoring of the health status of middle-aged men, unnecessary data collection and medical examinations are reduced, resource utilization efficiency is improved, and a closed-loop health management data flow that is dynamically adjusted according to health status is formed.

[0075] In a possible implementation, the method further includes: Step S210, after generating the personalized health management plan, extracting key feature activation paths from the spatiotemporal feature extraction network, and constructing a feature contribution heat map to visualize the degree of influence of different health data on risk prediction results.

[0076] After generating a personalized health management plan for middle-aged men, key feature activation paths can also be extracted from the spatiotemporal feature extraction network to construct a feature contribution heat map to visualize the degree of influence of different health data on risk prediction results. Taking a middle-aged man as an example, when the spatiotemporal feature extraction network processes his health data, blood pressure data in physiological monitoring data may be a key feature for predicting the probability of cardiovascular disease risk. When constructing a feature contribution heat map, the area corresponding to the blood pressure data may show a higher contribution value on the heat map, indicating that blood pressure has an important influence on the prediction results of cardiovascular disease risk. Similarly, the exercise time and sleep quality indicators in the behavioral activity record data also contribute to risk prediction. If his exercise time is short and his sleep quality is poor, the areas corresponding to these two features on the feature contribution heat map will show the corresponding contribution values, and different colors or values ​​are used to intuitively show the degree of influence of these health data on the risk prediction results.

[0077] Step S220, generating an execution effect simulation curve for each intervention item in the intervention measure priority list, showing the changing trend of the expected risk probability under different execution intensities.

[0078] For example, for the intervention item of dietary adjustment recommendation, assuming that the current daily salt intake of middle-aged men is 8 grams, the execution intensity is adjusted according to different salt intake levels, such as gradually reducing salt intake to different levels such as 6 grams, 5 grams, and 4 grams, and the expected risk probability will change accordingly. When salt intake is reduced to 6 grams, the risk probability of cardiovascular disease may be reduced from 30% to 28% due to the positive effect on blood pressure; when salt intake is further reduced to 5 grams, the risk probability may be reduced to 26%. Plotting the expected risk probability change data under these different execution intensities into an execution effect simulation curve can clearly show the potential effect of the dietary adjustment intervention item on reducing risk under different execution intensities. For the exercise prescription intervention item, such as the weekly increase in brisk walking exercise time from 30 minutes to 40 minutes, 50 minutes, etc., as the exercise time increases, the expected risk probability will also decrease accordingly, and the corresponding execution effect simulation curve is also drawn. For the medical examination plan intervention item, although it is mainly used to monitor health status, different examination frequencies may also affect the early detection and intervention timing of the disease, thereby indirectly affecting the risk probability, and the corresponding execution effect simulation curve is also generated for it.

[0079] Step S230, integrating the feature contribution heat map and the execution effect simulation curve into an interactive explanation report, and converting the visualization results into a popular health advice text for users through a natural language generation algorithm.

[0080] For example, after generating a personalized health management plan for middle-aged men, the feature contribution heat map and the execution effect simulation curve can be integrated into an interactive explanatory report. This interactive explanatory report not only contains intuitive visual graphics, but also uses natural language generation algorithms to convert these graphic information into popular health advice text that is easy for users to understand.

[0081] For example, in the feature contribution heat map, if it is shown that blood pressure data has the greatest impact on the prediction results of cardiovascular disease risk, then in the popular health advice text, it can be clearly stated: "Blood pressure data has a significant impact on your cardiovascular disease risk, so controlling blood pressure is the key to reducing risk." At the same time, combined with the execution effect simulation curve, if the salt intake adjustment in the dietary adjustment recommendation can significantly reduce the risk probability, the popular health advice text will further explain: "By gradually reducing daily salt intake, for example, from the current 8 grams to 6 grams, it is expected that your cardiovascular disease risk probability can be effectively reduced." In addition, the interactive interpretation report also allows users to interact with the graph, such as by clicking or sliding to view the expected risk probability change trend under different execution intensities. At the same time, the popular health advice text can be updated in real time based on the user's interaction to provide personalized health guidance and advice. In this way, users can not only intuitively understand how their health data affects disease risk, but also take corresponding health management measures based on the popular health advice text provided, thereby improving the effectiveness and pertinence of health management.

[0082] In a possible implementation, the method further includes: Step S310, embedding a feedback interface in the interactive interpretation report to receive the user's subjective execution difficulty score and physiological feeling description text of the health management plan through the feedback interface.

[0083] Step S320, converting the user feedback data into a feedback feature vector, concatenating it with the current health status representation vector and inputting it into a feedback adaptation network to generate a management plan adjustment coefficient.

[0084] Step S330: dynamically modify the order of items and execution intensity parameters in the intervention measure priority list according to the adjustment coefficient, and recalculate the dynamic monitoring cycle configuration parameters.

[0085] For example, a feedback interface is embedded in the interactive interpretation report, through which middle-aged men can provide feedback on the health management plan, including the subjective execution difficulty score and the physiological feeling description text. For example, a middle-aged man may find it difficult to increase the brisk walking time by 30 minutes per week according to the current exercise prescription. He gave a higher difficulty score in the subjective execution difficulty score, and mentioned in the physiological feeling description text that "I feel very tired after increasing exercise and it is difficult to persist." The user feedback data is converted into a feedback feature vector. This process involves quantification of the subjective execution difficulty score and the physiological feeling description text. For example, the subjective execution difficulty score is converted according to a certain numerical range, and a high difficulty score corresponds to a higher numerical value; for the physiological feeling description text, the key information is extracted through natural language processing technology, such as "fatigue" and "difficult to persist", and converted into quantifiable feature values, and then spliced ​​with the current health status representation vector.

[0086] The management scheme adjustment coefficient reflects the degree of adjustment of the health management scheme based on user feedback. The order of items and execution intensity parameters in the priority list of intervention measures are dynamically modified according to the management scheme adjustment coefficient, and the dynamic monitoring cycle configuration parameters are recalculated. For example, if a middle-aged male reports that the exercise prescription is difficult to execute and that he is physically tired, the adjustment coefficient may indicate that the execution intensity of the exercise prescription should be reduced. The original increase of 30 minutes of brisk walking time per week may be modified to an increase of 15 minutes of brisk walking time per week according to the adjustment coefficient. At the same time, due to the reduction in exercise intensity, the priority of dietary adjustment recommendations can be adjusted accordingly, and more attention can be paid to dietary adjustment to balance the overall health management effect. For the dynamic monitoring cycle configuration parameters, since the reduction in exercise intensity may affect the speed of improvement in health status, the comprehensive collection of physiological monitoring data and behavioral activity recording data, which was originally conducted every two weeks, may be adjusted to once every three weeks after recalculation to adapt to the new health management strategy. In this way, the health management plan is continuously optimized through user feedback to make it more in line with the actual situation and needs of middle-aged men, and to improve the effectiveness and feasibility of the health management plan.

[0087] In a possible implementation, the training process of the spatiotemporal feature extraction network includes: Step S101, constructing a training data set including target health data samples and their corresponding disease diagnosis labels, wherein each target health data sample includes a standardized health feature sequence of at least a continuous set time period.

[0088] In detail, for a target user group such as middle-aged men, each target health data sample contains a standardized health feature sequence for at least a continuous set time period. For example, samples are collected from the health data of many middle-aged men, and one sample may contain a standardized health feature sequence for a continuous month. The standardized health feature sequence covers physiological monitoring data (such as heart rate, blood pressure, etc.), behavioral activity record data (such as exercise status, sleep quality, etc.), and medical history text data (such as past disease diagnosis, symptom description, etc.), and each sample has a corresponding disease diagnosis label, such as whether it has cardiovascular disease, diabetes, etc.

[0089] Step S102, constructing a multi-task loss function, which includes time series reconstruction loss, cross-modal consistency loss and disease classification loss. The time series reconstruction loss requires the network to be able to reconstruct the original input sequence from high-level features, the cross-modal consistency loss constrains the distribution alignment of different modal features in the latent space, and the disease classification loss supervises the accuracy of risk prediction.

[0090] Taking the heart rate data of middle-aged men as an example, when the spatiotemporal feature extraction network processes the standardized health feature sequence containing the heart rate data and obtains high-level features, it should be able to restore the original heart rate data sequence as accurately as possible based on these high-level features. If the original heart rate data is 70 beats / minute at a certain moment, the value reconstructed by the spatiotemporal feature extraction network should be as close to this value as possible. The cross-modal consistency loss constrains the distribution alignment of different modal features in the latent space. For example, the distribution of the two different modal features of blood pressure in physiological monitoring data and exercise energy consumption in behavioral activity recording data should be as consistent as possible after being converted to the latent space by network processing, so as to ensure that the network can correctly understand and associate the relationship between different modalities when processing multimodal data. Disease classification loss supervises the accuracy of risk prediction. When the network processes the health data of middle-aged men and predicts whether they have a certain disease (such as cardiovascular disease), the prediction result should be as consistent as possible with the actual disease diagnosis label. If the person actually has cardiovascular disease and the network predicts that he is not ill, then the disease classification loss will be large.

[0091] Step S103, adopting a curriculum learning strategy to gradually increase the time span and modal complexity of the training data set, generating a pre-trained model of the spatiotemporal feature extraction network, using short-term data of a single modality in the initial training stage, and long-term continuous data of all modalities in the final stage.

[0092] In a possible implementation, step S103 specifically includes: Step S1031, obtaining an initial training stage data set corresponding to the training data set, wherein the initial training stage data set only contains short-term time series segments of a single health monitoring modality, wherein the continuous time length of the short-term time series segments does not exceed a preset initial time window threshold, and the single health monitoring modality is selected from one of the physiological monitoring data, behavioral activity record data, or medical history text data.

[0093] For example, physiological monitoring data can be selected as a single health monitoring modality, and the continuous time length of the short-term time series segment does not exceed the preset initial time window threshold. Assume that the preset initial time window threshold is one week. From the health data of many middle-aged men, only the time series segments containing physiological monitoring data (such as heart rate, blood pressure, etc.) within this week are selected to form the initial training stage data set.

[0094] Step S1032, generating a first training batch based on the initial training phase data set, inputting the first training batch into the spatiotemporal feature extraction network for feature extraction capability pre-training, and outputting an initial network weight parameter set, wherein the parameters of other network layers except the input layer are frozen during the pre-training process.

[0095] In the pre-training process, the parameters of other network layers except the input layer are frozen. The main purpose is to let the spatiotemporal feature extraction network learn the basic feature representation of single modality data. For example, the spatiotemporal feature extraction network first learns the basic fluctuation patterns and feature relationships of heart rate and blood pressure in physiological monitoring data, and outputs the initial network weight parameter set.

[0096] Step S1033, expanding the modality type of the data set in the initial training phase to two target modalities, and keeping the time length of the short-term time series segment unchanged, to generate a second training batch, wherein the combination of the two target modalities is selected from physiological monitoring data and behavioral activity recording data, physiological monitoring data and medical history text data, or behavioral activity recording data and medical history text data.

[0097] Step S1034, initializing the parameters of the spatiotemporal feature extraction network based on the initial network weight parameter set, using the second training batch to perform cross-modal association training, updating the network layer parameters related to the cross-modal attention mechanism in the spatiotemporal feature extraction network, and generating an intermediate network weight parameter set.

[0098] For example, physiological monitoring data and behavioral activity recording data are selected as two target modalities, and physiological monitoring data (such as heart rate, blood pressure, etc.) and behavioral activity recording data (such as exercise steps, sleep quality, etc.) within a week are combined to form the second training batch. The parameters of the spatiotemporal feature extraction network are initialized based on the initial network weight parameter set obtained previously, and the second training batch is used for cross-modal association training. In this process, the focus is to let the network learn the association relationship between the two different modal data, update the network layer parameters related to the cross-modal attention mechanism in the spatiotemporal feature extraction network, such as learning the potential association between heart rate and exercise steps, and the relationship between blood pressure and sleep quality, etc., to generate an intermediate network weight parameter set.

[0099] Step S1035, extending the time length of the second training batch to the mid-term time window threshold to generate a third training batch, wherein the mid-term time window threshold is a preset multiple of the initial time window threshold, and during the expansion process, time alignment and missing value interpolation are performed on the cross-modal data of the newly added time period.

[0100] Step S1036, based on the intermediate network weight parameter set, the spatiotemporal feature extraction network is loaded with parameters, the third training batch is used for long time series dependency training, the network layer parameters related to the temporal convolution operation in the spatiotemporal feature extraction network are updated, and an enhanced network weight parameter set is generated.

[0101] The time length of the second training batch is extended to the mid-term time window threshold. Assume that the mid-term time window threshold is three times the initial time window threshold (one week), that is, three weeks. During the extension process, the cross-modal data of the newly added time period is time-aligned and missing values ​​are interpolated. For example, for the newly added physiological monitoring data and behavioral activity recording data within two weeks, ensure that they are correctly aligned in time with the data of the previous week. If there are missing values ​​(such as missing exercise steps on a certain day), use appropriate interpolation methods (such as mean interpolation) to supplement them. Load the parameters of the spatiotemporal feature extraction network based on the intermediate network weight parameter set, and use this third training batch containing three weeks of data for long-term dependency training. In this training process, the network learns the feature change pattern under long time series, updates the network layer parameters related to the time convolution operation in the spatiotemporal feature extraction network, such as learning the change law of heart rate over time and the long-term trend of exercise steps in three weeks, and generates a set of enhanced network weight parameters.

[0102] Step S1037, expanding the modality type of the third training batch to a full modality type, and synchronously expanding its time length to a long-term time window threshold, to generate a fourth training batch, wherein the full modality type includes physiological monitoring data, behavioral activity record data, and medical history text data, and the long-term time window threshold covers a complete health monitoring cycle of at least a continuous preset number of days.

[0103] Step S1038, initializing the parameters of the spatiotemporal feature extraction network based on the enhanced network weight parameter set, using the fourth training batch to perform multimodal long time series joint training, synchronously updating all trainable parameters in the spatiotemporal feature extraction network, and generating a final network weight parameter set.

[0104] Step S1039, updating the architecture parameters of the spatiotemporal feature extraction network based on the final network weight parameter set to generate a pre-trained model suitable for multimodal long-time series health data analysis.

[0105] Assume that the long-term time window threshold covers a complete health monitoring cycle of at least a preset number of consecutive days (such as 30 consecutive days). The full-modal types include physiological monitoring data (such as heart rate, blood pressure, weight, etc.), behavioral activity record data (such as exercise status, sleep quality, etc.) and medical history text data (such as past disease diagnosis, symptom description, etc.). The spatiotemporal feature extraction network parameters are initialized based on the enhanced network weight parameter set, and the fourth training batch containing 30 days of full-modal data is used for multimodal long-term joint training. In this process, the network simultaneously learns the complex relationship between multimodal data and the law of feature changes under long time series, and synchronously updates all trainable parameters in the spatiotemporal feature extraction network to generate the final network weight parameter set. Finally, based on the final network weight parameter set, the architecture parameters of the spatiotemporal feature extraction network are updated to generate a pre-trained model adapted for multimodal long-term health data analysis. The pre-trained model can effectively process the multimodal long-term health data of target users such as middle-aged men.

[0106] Step S104, introducing healthy population data as negative samples through an adversarial training mechanism, enhancing the recognition sensitivity of the pre-trained model to abnormal health conditions, and generating the spatiotemporal feature extraction network.

[0107] For example, in a possible implementation, step S104 specifically includes: Step S1041, obtain an initial spatiotemporal feature extraction network parameter set after course learning strategy training, and an adversarial training data set including a healthy population data set and an abnormal health state sample set, wherein the samples in the healthy population data set are labeled as negative sample labels, and the samples in the abnormal health state sample set are labeled as positive sample labels.

[0108] First, obtain the initial spatiotemporal feature extraction network parameter set after training with the course learning strategy. This is the pre-training result obtained by gradually increasing the time span and modal complexity of the training data set. At the same time, prepare an adversarial training data set containing a healthy population data set and an abnormal health state sample set. Taking middle-aged men as an example, the middle-aged male samples in the healthy population data set have normal health indicators, such as a heart rate of 60-70 beats / minute, blood pressure of about 120 / 80 mmHg, normal exercise status (such as moderate exercise every day, weekly exercise time reaches a certain standard), and no medical history of major diseases. These samples are marked as negative sample labels, indicating a healthy state. The middle-aged male samples in the abnormal health state sample set may suffer from cardiovascular disease, their heart rate may often exceed 80 beats / minute, blood pressure is above 140 / 90 mmHg, their exercise state may be affected by the disease and unstable (such as decreased exercise endurance and easy fatigue), and there are diagnostic records of cardiovascular disease in the medical history. These samples are marked as positive sample labels, indicating a diseased state.

[0109] Step S1042, constructing an adversarial sample generator network, inputting the healthy population data set into the adversarial sample generator network, and generating a synthetic health data stream having a similar statistical distribution to the abnormal health status sample but marked as a negative sample label.

[0110] The adversarial sample generator network generates synthetic health data streams with similar statistical distributions to abnormal health state samples but labeled as negative samples by learning the statistical distribution characteristics of abnormal health state samples. For example, for the heart rate data in the healthy population data, although it is generally within the normal range, the adversarial sample generator network will generate some special cases of heart rate data based on the fluctuation pattern of the heart rate in the abnormal health state samples and the correlation with other physiological indicators, such as heart rate fluctuation data under the influence of some potential health risk factors, but these data are still within the normal range, thus forming a synthetic health data stream.

[0111] Step S1043, mixing the synthetic health data stream with the real abnormal health status sample set in a preset ratio to generate adversarial training batch data, and inputting the adversarial training batch data into the initial spatiotemporal feature extraction network for feature extraction to obtain an initial health status feature set.

[0112] Assume that the preset ratio is 1:1, that is, the number of samples in the synthetic health data stream is the same as the number of samples in the real abnormal health state sample set. Then the adversarial training batch data is input into the initial spatiotemporal feature extraction network for feature extraction to obtain the initial health state feature set. In this process, the spatiotemporal feature extraction network processes the mixed sample data and extracts the feature representation of each sample according to its existing network structure and parameters. These features cover the comprehensive features of multimodal information such as physiological monitoring data, behavioral activity record data, and medical history text data to form an initial health state feature set.

[0113] Step S1044, input the initial health status feature set into the adversarial discriminator network, calculate the discrimination loss value of the adversarial discriminator network for positive and negative samples, and back-propagate and update the generation parameters of the adversarial sample generator network and the discrimination parameters of the adversarial discriminator network according to the discrimination loss value.

[0114] If the adversarial discriminator network mistakenly judges the samples in the synthetic health data stream as positive samples (i.e., judged as abnormal health status), or judges the real abnormal health status samples as negative samples (i.e., judged as healthy status), a large discrimination loss value will be generated. According to the discrimination loss value, the generation parameters of the adversarial sample generator network and the discrimination parameters of the adversarial discriminator network are back-propagated to update. For example, if a sample in a synthetic health data stream is misjudged as abnormal health status, it means that the sample generated by the adversarial sample generator network deviates from the distribution of real healthy population data, and its generation parameters need to be adjusted to make the generated samples more consistent with the characteristics of healthy population data; at the same time, the discrimination parameters of the adversarial discriminator network also need to be adjusted to improve its ability to distinguish between positive and negative samples.

[0115] Step S1045, regenerate a new synthetic health data stream with the updated adversarial sample generator network, and iteratively perform the mixing, feature extraction and discrimination loss calculation process of the adversarial training batch data until the adversarial discriminator network cannot distinguish between real abnormal samples and synthetic health data with a preset accuracy threshold.

[0116] Assuming that the preset accuracy threshold is 90%, during the iteration process, as the adversarial sample generator network is continuously optimized, the generated synthetic health data stream is getting closer and closer to the real healthy population data distribution, and the adversarial discriminator network gradually becomes more difficult to distinguish between positive and negative samples. When the adversarial discriminator network's accuracy in distinguishing samples is close to 90%, it means that the synthetic health data stream and the real abnormal health status samples are already very similar in features and difficult to accurately distinguish.

[0117] Step S1046, freeze the parameters of the adversarial sample generator network, input the healthy population data set and the abnormal health status sample set into the adversarial sample generator network, and generate an adversarial enhanced feature set, wherein the adversarial enhanced feature set includes negative sample feature perturbation items and positive sample feature enhancement items.

[0118] In this process, since the parameters of the adversarial sample generator network are fixed, it will perform specific processing on the input healthy population data and abnormal health status samples to generate an adversarial enhanced feature set containing negative sample feature perturbation items and positive sample feature enhancement items. For healthy population data, such as normal blood pressure data, some tiny negative sample feature perturbation items may be added to make the representation of healthy population data in the feature space more diverse, which helps the adversarial sample generator network to better distinguish between healthy people and abnormal health status; for abnormal health status samples, such as hypertension data of middle-aged men with cardiovascular diseases, positive sample feature enhancement items will be added to highlight the difference in features with healthy population data, such as strengthening the abnormal correlation between blood pressure and other physiological indicators.

[0119] Step S1047, performing feature-level splicing on the adversarial enhancement feature set and the initial health status feature set to generate an adversarial training feature matrix, and inputting the adversarial training feature matrix into the spatiotemporal feature extraction network for feature reconstruction training.

[0120] The adversarial training feature matrix contains the comprehensive feature information of healthy people and abnormal health status samples after special processing. Then the adversarial training feature matrix is ​​input into the spatiotemporal feature extraction network for feature reconstruction training. The spatiotemporal feature extraction network attempts to reconstruct outputs similar to the original adversarial training batch data based on the adversarial training feature matrix. For example, for the health data of middle-aged men, if the original adversarial training batch data contains information such as heart rate, blood pressure, exercise status, and medical history at a specific time point, the spatiotemporal feature extraction network should reconstruct this information as accurately as possible.

[0121] Step S1048, calculating the distribution consistency loss between the reconstructed output of the spatiotemporal feature extraction network and the original adversarial training batch data, combining the feature confusion loss provided by the adversarial discriminator network, and generating a total adversarial training loss value.

[0122] Step S1049, performing gradient optimization on the parameters of the spatiotemporal feature extraction network according to the total loss value of the adversarial training, and updating the convolution kernel weights and attention mechanism parameters related to cross-modal feature extraction in the spatiotemporal feature extraction network.

[0123] If the distribution of data such as heart rate and blood pressure reconstructed by the spatiotemporal feature extraction network is significantly different from the distribution of the original data, a large distribution consistency loss will occur; at the same time, if the adversarial discriminator network has a high degree of confusion about the reconstructed features, the feature confusion loss will also increase. The two are added together to obtain the total loss value of adversarial training. According to the total loss value of adversarial training, the parameters of the spatiotemporal feature extraction network are gradient optimized, focusing on updating the convolution kernel weights and attention mechanism parameters related to cross-modal feature extraction in the spatiotemporal feature extraction network. For example, the convolution kernel weights are adjusted to better capture the relationship between different modal data, and the attention mechanism parameters are optimized to increase the focus on key features.

[0124] Step S10410, dynamically adjust the ratio of the synthetic healthy data stream to the real abnormal samples in the adversarial training batch data, so that during the training process, the ratio of the synthetic healthy data stream decreases with the increase of training rounds until it is completely replaced by the real abnormal healthy state sample set.

[0125] At the beginning, the ratio of the synthetic healthy data stream to the real abnormal samples may be 1:1. As the number of training rounds increases, it is gradually adjusted to 1:2, 1:3, etc., until the real abnormal health state sample set is used for training. In this process, adversarial sample generation, feature enhancement, discriminant loss calculation, and network parameter update operations are repeated until the inter-class distance in the latent space between the feature extraction results of the spatiotemporal feature extraction network for the healthy population data set and the feature extraction results of the abnormal health state sample set exceeds the preset threshold. Assuming the preset threshold is 0.8, this means that the spatiotemporal feature extraction network can well distinguish the feature representations of healthy people and abnormal health states in the latent space, that is, the inter-class distance between the two is large enough.

[0126] Step S10411, repeatedly perform adversarial sample generation, feature enhancement, discrimination loss calculation and network parameter update operations until the inter-class distance in the latent space between the feature extraction results of the spatiotemporal feature extraction network for the healthy population data set and the feature extraction results of the abnormal health status sample set exceeds a preset threshold.

[0127] Step S10412, decoupling the finally optimized spatiotemporal feature extraction network parameters from the parameters of the adversarial discriminator network, and only retaining the spatiotemporal feature extraction network parameters as an enhanced model for abnormal health status recognition.

[0128] Step S10413, verifying the classification accuracy of the enhanced model for healthy people and abnormal health conditions on an independent test set. When the classification accuracy does not reach a preset performance threshold, readjusting the perturbation intensity parameter of the adversarial sample generator network and iterating the adversarial training process.

[0129] The independent test set contains healthy data samples of middle-aged men who did not participate in the training, including both healthy people and abnormal health status samples. If the classification accuracy does not reach the preset performance threshold (for example, the preset is 95%), readjust the perturbation intensity parameter of the adversarial sample generator network and iterate the adversarial training process. For example, if the classification accuracy is 90%, appropriately increase the perturbation intensity of the adversarial sample generator network and perform the above adversarial training process again until the classification accuracy reaches or exceeds 95%.

[0130] Step S10414, performing parameter fusion on the verified enhanced model and the pre-trained model trained by the course learning strategy to generate a spatiotemporal feature extraction network with cross-modal abnormality sensitivity.

[0131] In this embodiment, the final spatiotemporal feature extraction network has both the ability to process multimodal long-term health data acquired through the curriculum learning strategy and the enhanced recognition sensitivity to abnormal health states through the adversarial training mechanism, thereby being able to more accurately process the health data of target users such as middle-aged men.

[0132] In a possible implementation, the method further includes: Step S410, after the training is completed, channel pruning and quantization-aware training are performed on the spatiotemporal feature extraction network to remove redundant feature channels and convert floating-point weights into fixed-point representations.

[0133] For example, in the spatiotemporal feature extraction network, some convolutional layers may have channels that contribute less to the representation of health data features. By analyzing the importance of each channel in processing middle-aged male health data, for example, observing its contribution to the extraction of key features such as heart rate and blood pressure, those channels with less contribution are removed. Quantization-aware training converts floating-point weights into fixed-point representation. When processing middle-aged male health data, the weights in the original network may be floating-point data. Quantization-aware training converts these weights into fixed-point numbers, reducing storage space and computation without losing too much precision, thereby improving the network's operating efficiency.

[0134] Step S420, construct a lightweight feature cache mechanism to store the preprocessing results of the standardized health feature sequence and the low-dimensional health status representation vector in the local memory of the edge computing device.

[0135] For the health data of middle-aged men, the standardized health feature sequence is preprocessed, such as the results of the noise filtering and feature alignment operations mentioned above, and the low-dimensional health status representation vector obtained by the spatiotemporal feature extraction network is stored in the local memory of the edge computing device (such as a wearable device or a local health monitoring device). In this way, when health risk prediction or health management plan generation is required, these data can be directly obtained from the local memory, reducing the time for data transmission and recalculation and improving the response speed of the system.

[0136] Step S430, construct a multi-threaded reasoning engine, decompose the calculation task of the health risk prediction model into subtasks executed in parallel, and realize the generation and update of real-time health management solutions.

[0137] Taking middle-aged men as an example, when the health risk prediction model calculates tasks such as cardiovascular disease risk probability, metabolic syndrome risk index, and mental stress assessment score, these tasks can be decomposed into multiple subtasks. For example, part of the task of calculating the cardiovascular disease risk probability can be performed simultaneously with part of the task of calculating the metabolic syndrome risk index. The multi-threaded reasoning engine reasonably allocates computing resources to enable these subtasks to be executed in parallel, thereby speeding up the calculation speed of the health risk prediction model, and can timely generate or update health management plans based on the latest health data of middle-aged men, such as adjusting the priority list of intervention measures or dynamically monitoring cycle configuration parameters, so as to achieve real-time monitoring and effective management of the health status of middle-aged men.

[0138] Figure 2 The hardware structure of the health management platform system 100 provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the health management platform system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0139] In one possible design, the health management platform system 100 can be a single server or a server group. The server group can be centralized or distributed (for example, the health management platform system 100 can be a distributed system). In some embodiments, the health management platform system 100 can be local or remote. For example, the health management platform system 100 can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the health management platform system 100 can be directly connected to a machine-readable storage medium 120 to access stored information and / or data. In some embodiments, the health management platform system 100 can be implemented on a health management platform system. By way of example only, the health management platform system may include a private cloud, a semantically related cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any aggregation thereof.

[0140] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions used by the health management platform system 100 to execute or use to complete the exemplary methods described in this application.

[0141] During the specific implementation process, one or more processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the health management data mining method based on deep learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0142] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned health management platform system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0143] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are set. When a processor executes the computer executable instructions, the above-mentioned health management data mining method based on deep learning is implemented.

[0144] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A health management data mining method based on deep learning, characterized in that: The method comprises: Acquire a target health data set of a target user, wherein the target health data set includes physiological monitoring data, behavioral activity record data, and medical history text data, wherein the physiological monitoring data includes multiple types of continuous physiological parameters, the behavioral activity record data includes movement status and sleep quality indicators in a time series, and the medical history text data includes structured diagnosis records and unstructured symptom descriptions; Performing cross-modal noise filtering and feature alignment processing on the target health data set to generate a standardized health feature sequence, wherein each feature unit in the standardized health feature sequence includes a synchronization timestamp, a cross-modal association identifier, and a normalized numerical expression; The standardized health feature sequence is hierarchically abstracted through a cascaded spatiotemporal feature extraction network to obtain a multi-granularity health state representation vector, wherein the spatiotemporal feature extraction network includes a temporal convolution branch and a spatial attention branch arranged in parallel, wherein the temporal convolution branch is used to capture long-term health trends, and the spatial attention branch is used to identify dynamic dependencies between cross-modal features; Inputting the multi-granularity health status representation vector into a pre-trained health risk prediction model to generate a personalized health management plan for the target user, wherein the personalized health management plan includes a disease risk level label, an intervention measure priority list, and a dynamic monitoring cycle configuration parameter; According to the dynamic monitoring cycle configuration parameters in the personalized health management solution, the collection frequency and feature alignment strategy of the target health data set are adjusted in real time to form a closed-loop health management data flow.

2. The health management data mining method based on deep learning according to claim 1 is characterized in that: The performing cross-modal noise filtering and feature alignment processing on the target health data set to generate a standardized health feature sequence specifically includes: Perform abnormal fluctuation detection on the continuous physiological parameters in the physiological monitoring data, use an adaptive threshold segmentation algorithm to identify the signal acquisition noise segment, and perform data repair on the signal acquisition noise segment through bidirectional cyclic interpolation to obtain the repaired physiological parameters; Classifying the motion state indicators in the behavior activity record data into motion types, converting discrete motion events into continuous energy consumption curves based on a preset motion energy consumption mapping table, and aligning the time windows with the sleep quality indicators to obtain an aligned behavior activity curve; Extracting entity relationships from the unstructured symptom descriptions in the medical history text data, constructing a symptom-diagnosis-treatment triple knowledge graph, and embedding the graph nodes in the triple knowledge graph into low-dimensional semantic vectors to obtain semantically vectorized medical knowledge; The repaired physiological parameters, aligned behavioral activity curves and semantically vectorized medical knowledge are subjected to multimodal feature fusion, and a dynamic time warping algorithm is used to eliminate the time offset of cross-modal data to generate the standardized health feature sequence with a unified time reference.

3. The health management data mining method based on deep learning according to claim 1 is characterized in that: The hierarchical feature abstraction of the standardized health feature sequence through the cascaded spatiotemporal feature extraction network to obtain a multi-granularity health status representation vector specifically includes: In the temporal convolution branch, a dilated causal convolution layer stacking structure is used to extract multi-scale temporal patterns from the standardized health feature sequence to obtain a multi-scale temporal feature graph, wherein convolution layers with different dilation rates capture short-term physiological fluctuations, medium-term behavioral patterns, and long-term health trends, respectively; In the spatial attention branch, a cross-modal feature association matrix is ​​constructed, the contribution of different modal features to the health status is calculated through a learnable attention weight allocation mechanism, and an inter-modal dependency graph is generated; The multi-scale temporal feature map output by the temporal convolution branch is concatenated with the inter-modal dependency graph generated by the spatial attention branch, and after the feature weights are adjusted by the gated fusion unit, they are input into the bidirectional gated recurrent network for context-aware feature enhancement to obtain enhanced fusion features; The enhanced fusion features are subjected to a hierarchical pooling operation to extract the global health status summary vector and the local abnormal pattern feature vector respectively, and are integrated through dimensionality reduction through a fully connected layer to generate the multi-granularity health status representation vector containing multi-level health information.

4. The health management data mining method based on deep learning according to claim 1 is characterized in that: The pre-trained health risk prediction model includes a multi-head risk prediction module and an intervention strategy generation module. The multi-granularity health state representation vector is input into the pre-trained health risk prediction model to generate a personalized health management plan for the target user, specifically including: In the multi-head risk prediction module, the cardiovascular disease risk probability, the metabolic syndrome risk index and the mental stress assessment score are calculated in parallel, wherein each risk prediction module uses an independent feature transformation layer and risk classifier and shares the underlying features of the multi-granularity health status representation vector; Determining the severity classification of the disease risk level label according to the weighted combination result of the cardiovascular disease risk probability and the metabolic syndrome risk index, and dynamically calibrating the classification result in combination with the mental stress assessment score; In the intervention strategy generation module, based on the calibrated disease risk level, matching candidate intervention items are retrieved from a preset intervention measure knowledge base, and the execution feasibility score of each intervention item is calculated based on the historical behavior activity record data of the target user; A reinforcement learning strategy is used to prioritize candidate intervention items, generate a priority list of intervention measures including dietary adjustment recommendations, exercise prescriptions, and medical examination plans, and derive the dynamic monitoring cycle configuration parameters based on the risk probability change rate.

5. The health management data mining method based on deep learning according to claim 1, characterized in that: The step of adjusting the acquisition frequency and feature alignment strategy of the target health data set in real time according to the dynamic monitoring cycle configuration parameters in the personalized health management solution to form a closed-loop health management data stream specifically includes: When it is detected that the severity level of the disease risk level label exceeds a preset threshold, the sampling interval of the physiological monitoring data is shortened to a preset ratio of the original sampling interval, and the number of activations of the motion posture sensor in the behavior activity record data is increased according to the set frequency increase strategy; Dynamically updating the standardized health feature sequence according to the multimodal data stream collected in real time, using a sliding window mechanism to perform incremental feature alignment processing on the newly added data segments, and retaining the contextual relevance of the historical feature sequence; Inputting the updated standardized health feature sequence into the spatiotemporal feature extraction network for online feature abstraction, and triggering the real-time reasoning process of the health risk prediction model to generate an updated personalized health management plan; When the disease risk level label remains stable for a preset number of consecutive monitoring cycles, the data collection interval is gradually extended until the baseline frequency is restored, while the intensity of the medical examination plan in the priority list of the intervention measures is reduced.

6. The health management data mining method based on deep learning according to claim 1, characterized in that: The method further comprises: After generating the personalized health management plan, extracting key feature activation paths from the spatiotemporal feature extraction network, and constructing a feature contribution heat map to visualize the degree of influence of different health data on risk prediction results; Generate an execution effect simulation curve for each intervention item in the priority list of intervention measures to show the changing trend of expected risk probability under different execution intensities; The feature contribution heat map and the execution effect simulation curve are integrated into an interactive explanation report, and the visualization results are converted into popular health advice text for users through a natural language generation algorithm.

7. The health management data mining method based on deep learning according to claim 6 is characterized in that: The method further comprises: Embedding a feedback interface in the interactive interpretation report to receive a user's subjective execution difficulty score and physiological feeling description text of the health management plan through the feedback interface; The user feedback data is converted into a feedback feature vector, which is then concatenated with the current health status representation vector and input into the feedback adaptation network to generate the management plan adjustment coefficient; The order of items and execution intensity parameters in the intervention measure priority list are dynamically modified according to the adjustment coefficient, and the dynamic monitoring cycle configuration parameters are recalculated.

8. The health management data mining method based on deep learning according to claim 1, characterized in that: The training process of the spatiotemporal feature extraction network includes: Constructing a training data set including target health data samples and their corresponding disease diagnosis labels, wherein each target health data sample includes a standardized health feature sequence for at least a set continuous time period; Constructing a multi-task loss function, the multi-task loss function includes time series reconstruction loss, cross-modal consistency loss and disease classification loss. The time series reconstruction loss requires the network to be able to reconstruct the original input sequence from high-level features, the cross-modal consistency loss constrains the distribution alignment of different modal features in the latent space, and the disease classification loss supervises the accuracy of risk prediction; A curriculum learning strategy is adopted to gradually increase the time span and modal complexity of the training data set to generate a pre-trained model of the spatiotemporal feature extraction network, where the initial training phase uses short-term data of a single modality and the final phase uses long-term continuous data of all modalities; Healthy population data is introduced as negative samples through an adversarial training mechanism to enhance the recognition sensitivity of the pre-trained model to abnormal health conditions and generate the spatiotemporal feature extraction network.

9. The health management data mining method based on deep learning according to claim 8, characterized in that: The adopting of the curriculum learning strategy to gradually increase the time span and modal complexity of the training data set to generate a pre-trained model of the spatiotemporal feature extraction network specifically includes: Acquire an initial training phase data set corresponding to the training data set, wherein the initial training phase data set only contains short-term time series segments of a single health monitoring modality, wherein the continuous time length of the short-term time series segments does not exceed a preset initial time window threshold, and the single health monitoring modality is selected from one of the physiological monitoring data, the behavioral activity record data, or the medical history text data; Generate a first training batch based on the data set in the initial training phase, input the first training batch into the spatiotemporal feature extraction network for feature extraction capability pre-training, and output an initial network weight parameter set, wherein parameters of other network layers except the input layer are frozen during the pre-training process; Expanding the modality type of the data set in the initial training phase to two target modalities and keeping the time length of the short-term time series segment unchanged to generate a second training batch, wherein the combination of the two target modalities is selected from physiological monitoring data and behavioral activity recording data, physiological monitoring data and medical history text data, or behavioral activity recording data and medical history text data; Initializing parameters of the spatiotemporal feature extraction network based on the initial network weight parameter set, performing cross-modal association training using the second training batch, updating network layer parameters related to the cross-modal attention mechanism in the spatiotemporal feature extraction network, and generating an intermediate network weight parameter set; Extending the time length of the second training batch to a mid-term time window threshold to generate a third training batch, wherein the mid-term time window threshold is a preset multiple of the initial time window threshold, and performing time alignment and missing value interpolation on the cross-modal data of the newly added time period during the extension process; Based on the intermediate network weight parameter set, the spatiotemporal feature extraction network is loaded with parameters, the third training batch is used to perform long time series dependency training, the network layer parameters related to the temporal convolution operation in the spatiotemporal feature extraction network are updated, and a reinforced network weight parameter set is generated; Expanding the modality type of the third training batch to a full modality type, and simultaneously expanding its time length to a long-term time window threshold, to generate a fourth training batch, wherein the full modality type includes physiological monitoring data, behavioral activity record data, and medical history text data, and the long-term time window threshold covers a complete health monitoring cycle of at least a continuous preset number of days; Initializing parameters of the spatiotemporal feature extraction network based on the enhanced network weight parameter set, using the fourth training batch to perform multimodal long time series joint training, synchronously updating all trainable parameters in the spatiotemporal feature extraction network, and generating a final network weight parameter set; Based on the final network weight parameter set, the architecture parameters of the spatiotemporal feature extraction network are updated to generate a pre-trained model suitable for multimodal long-time series health data analysis.

10. A health management platform system, characterized in that: The health management platform system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the deep learning-based health management data mining method described in any one of claims 1 to 9.

Images