Health management data mining method and system based on deep learning

Abstract

The application provides a health management data mining method and system based on deep learning. First, a target health data set of a target user is obtained, which includes physiological monitoring, behavior activity records, and medical history texts. Then, cross-modal noise filtering and feature alignment processing are performed on the target health data set to generate a standardized health feature sequence containing a synchronization timestamp. Next, a cascaded spatiotemporal feature extraction network is used to perform hierarchical feature abstraction using a parallel time series convolution branch and a spatial attention branch to obtain a multi-granularity health state representation vector. Then, the multi-granularity health state representation vector is input into a pre-trained model to generate an individualized health management plan containing disease risk levels and the like. Finally, the data acquisition frequency and feature alignment strategy are adjusted in real time according to the dynamic monitoring period parameters in the individualized health management plan to form a closed-loop health management data flow, thereby achieving precise and individualized health management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a health management data mining method and system based on deep learning. BACKGROUND

[0002] With the increasing attention to health, the field of health management has developed rapidly. In the existing health management technology, there are many limitations, which make the effect and accuracy of health management difficult to meet the growing demand.

[0003] Early health management mainly relies on manual recording and simple data statistical analysis. Medical staff manually collects patients' basic health information such as height, weight, blood pressure and other single-dimensional data, and then preliminarily evaluates the health status of patients through experience and simple calculation formula. This method is not only inefficient, but also difficult to accurately grasp the overall health status of patients due to the incompleteness of data collection and the simplicity of analysis method, and cannot discover potential health risks in advance.

[0004] With the development of information technology, the existing information technology has limited processing capacity for multi-source heterogeneous data, so that the potential value between data cannot be fully mined, 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 pre-set rules, lacking consideration of personalized features of different users. For users with different living habits, genetic background and health status, the same risk prediction model and management scheme are used, resulting in inaccurate prediction results and lack of targeted management scheme, which cannot truly meet the personalized health management needs of users.

[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 adjust flexibly according to the real-time changes of the user's health status. For example, for users with stable health status, data is still collected at a higher frequency, which not only wastes resources but also increases the burden on users; for users with fluctuating health status, if the data collection frequency and management strategy cannot be adjusted in time, the best intervention opportunity may be missed, affecting the effect of health management. SUMMARY

[0007] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a health management data mining method based on deep learning, which comprises:

[0008] obtain a target health data set of a target user, the target health data set comprising physiological monitoring data, behavior activity record data, and medical history text data, wherein the physiological monitoring data contains multiple types of continuous physiological parameters, the behavior activity record data contains motion state and sleep quality indicators in a time sequence, and the medical history text data contains structured diagnosis records and unstructured symptom descriptions;

[0009] perform 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 contains a synchronous timestamp, a cross-modal association identifier, and a normalized numerical representation;

[0010] obtain a target health data set of a target user, the target health data set comprising physiological monitoring data, behavior activity record data, and medical history text data, wherein the physiological monitoring data contains multiple types of continuous physiological parameters, the behavior activity record data contains motion state and sleep quality indicators in a time sequence, and the medical history text data contains structured diagnosis records and unstructured symptom descriptions;

[0011] input the multi-granularity health state representation vector into a pre-trained health risk prediction model to generate an individualized health management scheme for the target user, the individualized health management scheme containing a disease risk level label, an intervention measure priority list, and a dynamic monitoring period configuration parameter;

[0012] adjust the collection frequency and feature alignment strategy of the target health data set in real time according to the dynamic monitoring period configuration parameter in the individualized health management scheme to form a closed-loop health management data flow.

[0013] In still another aspect, the embodiments of the present application also provide a health management platform system, comprising a processor, a machine readable storage medium, the machine readable storage medium being connected with the processor, the machine readable storage medium being used for storing programs, instructions or codes, and the processor being used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0014] Based on the above aspects, the embodiment of the present application comprehensively collects a multi-source heterogeneous target health data set containing physiological monitoring data, behavior activity record data and medical history text data of a target user, and performs cross-modal noise filtering and feature alignment processing based on the multi-source heterogeneous target health data set, to generate a standardized health feature sequence. Each feature unit contains a synchronous timestamp, a cross-modal correlation identifier and a normalized numerical expression. The cross-modal correlation identifier closely correlates the data of different modalities, and the synchronous timestamp further enhances the consistency of the data in the time dimension. Furthermore, the parallelly arranged time sequence convolution branch and spatial attention branch work cooperatively, capture long-period health trends and identify dynamic dependency relationships between cross-modal features, respectively, can deeply mine data features in different dimensions, grasp the long-term change law of health data from the time dimension, analyze the complex interaction relationship between different modalities of data from the spatial dimension, and thus obtain a multi-granularity health state representation vector. Then, the personalized health management scheme generated based on the multi-granularity health state representation vector contains a disease risk level label, an intervention measure priority list and a dynamic monitoring period configuration parameter. The disease risk level label helps the user clearly understand the health risk condition of the user, the intervention measure priority list provides a clear health management focus and direction for the user and medical staff, and the dynamic monitoring period configuration parameter dynamically adjusts the monitoring period, so that the data acquisition frequency and feature alignment strategy can be flexibly and intelligently adjusted according to the real-time health state change of the user, to form a closed-loop health management data flow, thereby effectively reducing the data acquisition cost and resource consumption while ensuring the health management effect. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is an execution flow diagram of the health management data mining method based on deep learning provided by the embodiment of the present application.

[0016] Figure 2 is a hardware architecture diagram of the health management platform system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0017] The present application will be described in detail below with reference to the accompanying drawings of the specification, Figure 1 is a flow diagram of the health management data mining method based on deep learning provided by an embodiment of the present application, and the health management data mining method based on deep learning will be described in detail below.

[0018] In step S110, a target health data set of a target user is obtained, which includes physiological monitoring data, behavior activity record data, and medical history text data. The physiological monitoring data contains multiple types of continuous physiological parameters, the behavior activity record data contains motion state and sleep quality indicators in time series, and the medical history text data contains structured diagnosis records and unstructured symptom descriptions.

[0019] In this embodiment, the target health data set of a middle-aged male is taken as an example.

[0020] In detail, for the physiological monitoring data, multiple types of continuous physiological parameters can be continuously collected by wearable devices such as smart bands and smart body fat scales. For example, a smart band can monitor heart rate in real time. In a day, his heart rate may be relatively stable at about 60-70 times per minute when he gets up in the morning, rise to 100-120 times per minute when he exercises (such as brisk walking or climbing stairs), and fall back to a lower level when he rests at night. A smart body fat scale can measure parameters such as body weight, body fat rate, and muscle mass. His body weight may be about 75 kg, and the body fat rate is 25%. The above continuous physiological parameters constitute part of the physiological monitoring data.

[0021] In terms of behavior activity record data, motion state is recorded by a motion sensor in a mobile phone. It is assumed that he has a brisk walking exercise time every morning. The motion sensor can record his step count, step frequency, and motion speed, and other motion state information. During sleep, a smart mattress or a smart band can monitor sleep quality indicators such as sleep duration, deep sleep duration, light sleep duration, and the number of turns during sleep. For example, his sleep duration is about 7-8 hours every night, and the deep sleep duration is about 2-3 hours.

[0022] The medical history text data contains structured diagnosis records and unstructured symptom descriptions. For example, he has had a physical examination and disease diagnosis in a hospital before. The structured diagnosis records may show that he has been diagnosed with mild hypertension, with a blood pressure value of about 140 / 90 mmHg. The unstructured symptom description may include occasional dizziness symptoms, especially when standing for a long time or suddenly standing up, which makes him feel dizzy. The above information is recorded in the medical history text data. Thus, by obtaining data from these different channels, a complete target health data set of the target user is generated.

[0023] In step S120, the target health data set is processed by cross-modal noise filtering and feature alignment to generate a standardized health feature sequence, wherein each feature unit in the standardized health feature sequence contains a synchronous timestamp, a cross-modal association identifier, and a normalized numerical representation.

[0024] Continuing with the above example of a middle-aged male, when processing the target health data set, first, abnormal fluctuation detection is performed on the continuous physiological parameters in the physiological monitoring data. For example, due to the smart bracelet being affected by external electromagnetic interference or not being worn tightly during signal acquisition, the heart rate data occasionally appears abnormal fluctuations. 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 10 times / minute up and down in a calm state, when the heart rate suddenly jumps to 150 times / minute and the duration is very short, the algorithm identifies it as a noise segment. Then, the data repair is performed on the signal acquisition noise segment by bidirectional circular interpolation, so that the repaired heart rate data is more smooth and accurate.

[0025] For the motion state indicators in the behavior activity record data, the action type classification is performed, and his morning brisk walking is classified as an aerobic fitness type action. Then, based on the preset motion energy consumption mapping table, the brisk walking discrete action event is converted into a continuous energy consumption curve. Assuming that the energy consumption of brisk walking is related to step frequency and speed, the energy consumption value of each minute is obtained by the existing calculation formula, thereby forming a continuous energy consumption curve. Then, the energy consumption curve is aligned with the sleep quality indicators in a time window, for example, 24 hours as a time window, so that the motion state and sleep quality are considered in the same time frame, and the aligned behavior activity curve is obtained.

[0026] When processing the unstructured symptom description in the medical history text data, entity relationship extraction is performed. For his occasional dizziness symptom, the relationship between dizziness and mild hypertension is extracted, and a triple graph knowledge graph of symptom-diagnosis-treatment is constructed. For example, the triple relationship of dizziness (symptom) - mild hypertension (diagnosis) - drug treatment (treatment). Then, the graph nodes in the triple graph knowledge graph are embedded into low-dimensional semantic vectors, and the concepts such as dizziness and mild hypertension are converted into semantic vectorized medical knowledge that can be processed by computers.

[0027] Finally, the repaired physiological parameters, the aligned activity curves, and the semantically vectorized medical knowledge are fused to obtain multi-modal features. Since the data from different modalities can have a time offset, for example, the physiological monitoring data is continuously collected in real time, while the medical history text data is recorded in a certain time period in the past, a dynamic time warping algorithm is used to eliminate the time offset of the cross-modal data, and a standardized health feature sequence with a unified time reference is generated. Each feature unit contains a synchronized timestamp (e.g., August 1, 2024, 9:00 am), a cross-modal correlation identifier (used to identify which modality the feature comes from, such as physiological monitoring, activity, or medical history), and a normalized numerical representation (the data in different ranges is normalized to a value between 0 and 1, such as heart rate).

[0028] In step S130, the standardized health feature sequence is subjected to hierarchical feature abstraction by a cascaded spatio-temporal feature extraction network to obtain a multi-granularity health state representation vector. The spatio-temporal feature extraction network includes a parallelly arranged time series convolution branch and a spatial attention branch. The time series convolution branch is used to capture long-term health trends, and the spatial attention branch is used to identify dynamic dependency relationships between cross-modal features.

[0029] Still taking the standardized health feature sequence of the middle-aged man as an example, it is input into the spatio-temporal feature extraction network.

[0030] 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. Convolution layers with different dilation rates play different roles. For example, convolution layers with small dilation rates can capture short-term physiological fluctuations, such as the rapid rise and then rapid decline of his heart rate after a short period of exercise; convolution layers with moderate dilation rates are used to capture medium-term behavior patterns, such as the trend of his exercise time and sleep quality in the past week; and convolution layers with large dilation rates can capture long-term health trends, such as the overall trend of his weight and blood pressure over the past few months, thereby obtaining multi-scale time series feature maps.

[0031] In the spatial attention branch, a cross-modal feature correlation matrix is constructed. Taking his heart rate and blood pressure in the physiological monitoring data, exercise energy consumption in the activity record data, and mild hypertension diagnosis in the medical history text data as examples, the contribution of these different modalities to the health state is calculated through a learnable attention weight distribution mechanism. For example, heart rate and blood pressure may have a high weight on his health state because he has mild hypertension, and exercise energy consumption is also associated with controlling blood pressure and improving overall health status. A cross-modal dependency graph is generated by calculating the contribution of each modality to the health state.

[0032] The multi-scale time series feature maps output by the time convolution branch are concatenated with the inter-modal dependency graphs generated by the spatial attention branch. Assuming that the multi-scale time series feature maps contain heart rate, blood pressure, movement, and sleep feature information at different time scales, and the inter-modal dependency graphs contain the correlation weight information between modal features, the two are concatenated and then adjusted by a gating fusion unit to adjust the feature weights. For example, if the blood pressure is more critical to the health status at the current stage, the feature weight related to the blood pressure will be appropriately increased during fusion. Then, the adjusted features are input into a bidirectional gated recurrent network for context-aware feature enhancement, enabling the network to better utilize the context information to understand the health status features. After obtaining the enhanced fusion features, hierarchical pooling operations are performed. First, a global health status summary vector is extracted, which can reflect the overall health status, such as the overall health risk level being moderate; at the same time, a local abnormal pattern feature vector is extracted, which can highlight the occasional dizziness of the individual. Finally, a dimension reduction integration is performed through a fully connected layer to generate a multi-granularity health status representation vector containing multi-level health information.

[0033] In step S140, the multi-granularity health status representation vector is input into a pre-trained health risk prediction model to generate an individualized health management plan for the target user, which includes a disease risk level label, an intervention measure priority list, and dynamic monitoring cycle configuration parameters.

[0034] In detail, the multi-granularity health status representation vector of the middle-aged man described above can be input into a pre-trained health risk prediction model.

[0035] 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, due to his mild hypertension, the cardiovascular disease risk probability is relatively high, and the calculated probability is 30%. For the metabolic syndrome risk index, considering his weight, body fat rate, and other factors, the calculated risk index may be at a moderate level, such as 40%. In terms of mental stress assessment, by analyzing his behavior activity records (such as regular exercise and adequate sleep) and symptom descriptions in the 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 a risk classifier, but shares the underlying features of the multi-granularity health status representation vector.

[0036] The severity classification of the disease risk level label is determined according to the weighted combination result 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 and the weight of the metabolic syndrome risk index is 0.4, the weighted combination result is 34%, and the severity classification result is dynamically calibrated in combination with the mental stress assessment score, it is determined that his disease risk level is medium risk.

[0037] In the intervention strategy generation module, based on the calibrated disease risk level, the matched candidate intervention items are retrieved from the preset intervention measure knowledge base. Since he is at medium risk and has mild hypertension, the candidate intervention items may include dietary adjustment (such as reducing salt intake), exercise prescription (increasing aerobic exercise time and intensity), and medical examination plan (regularly measuring blood pressure, conducting blood lipid examination, etc.). Then the execution feasibility score of each intervention item is calculated according to his historical behavior activity record data. For example, he has certain exercise habits before, so the execution feasibility score of the exercise prescription may be higher. The reinforcement learning strategy is used to prioritize the candidate intervention items, and an intervention measure priority list containing dietary adjustment suggestions (such as daily salt intake of no more than 6 grams), exercise prescription (increasing 30 minutes of brisk walking exercise time per week), and medical examination plan (comprehensive physical examination every three months) is generated. At the same time, based on the risk probability change rate, the dynamic monitoring cycle configuration parameters are derived, and since his risk is at a medium level and is relatively stable, the dynamic monitoring cycle is configured to collect his physiological monitoring data and behavior activity record data once every two weeks, and to update the medical history data once every three months.

[0038] Step S150, according to the dynamic monitoring cycle configuration parameters in the individualized health management scheme, 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.

[0039] In detail, according to the individualized health management scheme formulated for middle-aged men, when data collection is performed according to the dynamic monitoring cycle, the severity classification of the disease risk level label is adjusted.

[0040] When his disease risk level is at medium risk, physiological monitoring data and behavioral activity record data are collected in a bi-weekly cycle. For example, smart bands and smart scales collect data at set time intervals, and motion sensors continuously record motion states during his daily activities. If the severity classification of his disease risk level label exceeds the preset threshold (assuming the preset threshold is 50%) during a certain stage, for example, due to a sudden increase in his blood pressure or the appearance of new symptoms such as palpitations, the sampling interval of physiological monitoring data is shortened to a preset proportion of the original sampling interval (e.g., to half of the original). Instead of collecting comprehensive physiological data every two weeks, it is now collected every week. At the same time, the number of motion posture sensor activations in the behavioral activity record data is increased according to the set frequency increase strategy, for example, from recording motion states every day to recording every half day, in order to more timely obtain his changes in exercise and health status.

[0041] The standardized health feature sequence is dynamically updated according to the real-time collected multi-modal data stream. When new data arrives, a sliding window mechanism is used to perform incremental feature alignment processing on the new data segment. For example, a week of newly collected multi-modal data is aligned with the previous historical data, and on the basis of maintaining the contextual relevance of the historical feature sequence, the new multi-modal data is integrated into the standardized health feature sequence.

[0042] The updated standardized health feature sequence is input into the spatio-temporal feature extraction network for online feature abstraction, and triggers the real-time inference process of the health risk prediction model, generating an updated personalized health management plan. If his health condition improves after a period of adjustment and the disease risk level decreases, when the disease risk level label remains stable for a continuous preset number of monitoring cycles (assuming three cycles), the data collection interval is gradually extended until it returns to the baseline frequency. For example, physiological data is collected every week and gradually returns to every two weeks, and the intensity of medical examination plans in the intervention measure priority list is reduced, such as from a comprehensive physical examination every three months to every six months. In this way, a closed-loop health management data flow is formed, which adjusts data collection and management strategies in real time according to the user's health status, and continuously optimizes personalized health management plans.

[0043] Based on the above steps, the embodiments of the present application comprehensively collect a multi-source heterogeneous target health data set containing physiological monitoring data, behavior activity record data and medical history text data of the target user, and perform cross-modal noise filtering and feature alignment processing based thereon to generate a standardized health feature sequence, each feature unit containing a synchronous timestamp, a cross-modal association identifier and a normalized numerical expression. The cross-modal association identifier closely associates data of different modalities, and the synchronous timestamp further enhances the consistency of the data in the time dimension. Furthermore, the parallelly arranged time convolution branch and spatial attention branch work cooperatively, respectively capturing long-period health trends and identifying dynamic dependency relationships between cross-modal features, enabling deep mining of data features in different dimensions, grasping long-term change laws of health data from the time dimension, and analyzing interaction relationships between different modalities of data from the space dimension, thereby obtaining a multi-granularity health state representation vector. Then, the personalized health management scheme generated based on the multi-granularity health state representation vector contains a disease risk level label, an intervention measure priority list and a dynamic monitoring period configuration parameter, the disease risk level label helping the user clearly understand the health risk condition of the user, the intervention measure priority list providing clear health management focus and direction for the user and medical staff, and the dynamic monitoring period configuration parameter dynamically adjusting the monitoring period, which can flexibly and intelligently adjust the data acquisition frequency and feature alignment strategy according to the real-time health state change of the user, form a closed-loop health management data flow, and effectively reduce the data acquisition cost and resource consumption while ensuring the health management effect.

[0044] In a possible implementation, step S120 specifically includes:

[0045] Step S121, detecting abnormal fluctuations in the continuous physiological parameters in the physiological monitoring data, using an adaptive threshold segmentation algorithm to identify signal acquisition noise segments, and performing data repair on the signal acquisition noise segments through bidirectional circular interpolation to obtain repaired physiological parameters.

[0046] In detail, for the continuous physiological parameters in the physiological monitoring data of the middle-aged man in the above example, such as heart rate data, there may be abnormal fluctuations due to device or external factors. Normally, the heart rate of the middle-aged man in a quiet state is in the range of 60-70 times / minute, and can rise to 80-90 times / minute when performing daily activities such as walking, and can reach 100-120 times / minute when the movement 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 a quiet state is set to 5 times / minute up and down. When the heart rate data collected at a certain time is 150 times / minute and the surrounding data is in the normal fluctuation range, the adaptive threshold segmentation algorithm determines that the data segment is a signal acquisition noise segment. Then, the data in the signal acquisition noise segment is repaired by bidirectional circular interpolation, for example, based on the normal heart rate data before and after the noise segment, the data is repaired according to a certain interpolation algorithm, so that the repaired data and the data before and after it are smoothly connected, and the repaired physiological parameters are obtained.

[0047] In step S122, the motion state indicator in the behavior activity record data is classified into action types, discrete action events are converted into continuous energy consumption curves based on a preset motion energy consumption mapping table, and the sleep quality indicator is aligned in a time window to obtain an aligned behavior activity curve.

[0048] In terms of behavior activity record data, for the morning fast walking of the middle-aged man, the motion state indicator needs to be classified into action types, which is classified as an aerobic fitness action. Based on the preset motion energy consumption mapping table, which contains the relationship between different step frequencies, speeds and energy consumption, the energy consumption value is calculated according to the step frequency and speed data during his fast walking. Assuming that his step frequency is 120 steps / minute and his speed is 4 km / h during fast walking, the energy consumption value per minute is calculated by the formula in the mapping table, and these discrete energy consumption values are converted into a continuous energy consumption curve over time. At the same time, his sleep quality indicator contains sleep duration, deep sleep duration and other information, and the energy consumption curve is aligned with the sleep quality indicator in a 24-hour time window. For example, in this 24-hour time window, the energy consumption peak value corresponding to the morning fast walking is aligned with the sleep duration and other indicators in the evening in time to obtain an aligned behavior activity curve.

[0049] In step S123, the unstructured symptom description in the medical history text data is subjected to entity relationship extraction, a triple knowledge graph of symptom-diagnosis-treatment is constructed, and the nodes in the triple knowledge graph are embedded into low-dimensional semantic vectors to obtain semantic vectorized medical knowledge.

[0050] The entity relationship extraction is performed on the unstructured symptom description in the medical history text data of middle-aged men, such as the symptom of occasional dizziness. For example, he is diagnosed with mild hypertension, and a triple graph knowledge graph of symptom-diagnosis-treatment, i.e. dizziness (symptom)-mild hypertension (diagnosis)-drug treatment (treatment), is constructed. For the graph nodes in the triple graph knowledge graph, such as the concepts of dizziness, mild hypertension and drug treatment, 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 semantic vectorized medical knowledge that can be processed by a computer. The low-dimensional semantic vector contains the representation of the semantic information of the dizziness symptom in the low-dimensional space, so that the computer can understand and process the relationship between these medical knowledge and other modal data.

[0051] In step S124, the repaired physiological parameters, the aligned behavior activity curve and the semantic vectorized medical knowledge are fused for multi-modal feature fusion, a dynamic time warping algorithm is used to eliminate the time offset of cross-modal data, and the standardized health feature sequence with a unified time reference is generated.

[0052] In detail, the physiological monitoring data of middle-aged men is continuously collected in real time, and the medical history text data is a record in the past, and there is a difference in time. Therefore, a dynamic time warping algorithm is needed to eliminate the time offset of cross-modal data, so that all data are under a unified time reference. For example, the mild hypertension diagnosis 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, wherein each feature unit contains a synchronous time stamp (such as August 1, 2024, 9:00 am), a cross-modal correlation identifier (identifying that the feature comes from physiological monitoring, behavior activity or medical history) and a normalized numerical expression (the numerical value after normalization of data in different ranges, such as heart rate normalized to 0-1).

[0053] In one possible implementation, step S130 specifically includes:

[0054] In step S131, in the time series convolution branch, a multi-scale time series pattern extraction is performed on the standardized health feature sequence by using a dilated causal convolution layer stacking structure to obtain a multi-scale time series feature map, wherein the convolution layers with different dilation rates capture short-term physiological fluctuations, medium-term behavior patterns and long-term health trends, respectively.

[0055] In detail, the standardized health feature sequence of the middle-aged man can be input into the time series convolution branch of the spatio-temporal feature extraction network, and a dilated causal convolution layer stacking structure is used to extract multi-scale time series patterns. Convolution layers with different dilation rates play different roles. Convolution layers with smaller dilation rates can capture short-term physiological fluctuations, such as the rapid increase and then the short-term decrease in heart rate after the middle-aged man performs short-time stair climbing activities. This short-term heart rate fluctuation can be accurately detected by convolution layers with smaller dilation rates. Convolution layers with moderate dilation rates can capture medium-term behavior patterns, such as the change trend of his exercise time and sleep quality in the past week, such as this week he reduced his exercise time by 20% and his sleep quality also decreased due to his busy work. This medium-term behavior pattern change can be captured by the convolution layer. Convolution layers with larger dilation rates are used to capture long-term health trends, such as the overall change 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 shown a slight upward trend. In this way, multi-scale time series feature maps are obtained.

[0056] In step S132, in the spatial attention branch, a cross-modal feature correlation matrix is constructed, the contribution of different modal features to the health state is calculated through a learnable attention weight distribution mechanism, and an inter-modal dependency graph is generated.

[0057] Taking the heart rate and blood pressure in the physiological monitoring data of the middle-aged man, the exercise energy consumption in the behavior activity record data, and the mild hypertension diagnosis in the medical history text data as examples, the contribution of these different modal features to the health state is calculated through a learnable attention weight distribution mechanism. Since he has mild hypertension, the heart rate and blood pressure have a relatively high weight on his health state, for example, the weight of heart rate is 0.3, the weight of blood pressure is 0.4, and the exercise energy consumption also has a certain correlation with his overall health state, and the weight may be 0.2. The weight of the mild hypertension diagnosis in the medical history is 0.1. Through such calculation, an inter-modal dependency graph is generated, which can clearly show the dependency relationship between each modal feature and the importance of the health state.

[0058] In step S133, the multi-scale time series feature maps output by the time series convolution branch are spliced with the inter-modal dependency graph generated by the spatial attention branch, and after adjusting the feature weights by the gating fusion unit, the enhanced fusion features are obtained by inputting them into the bidirectional gated recurrent network for context-aware feature enhancement.

[0059] Assuming that the multi-scale time sequence feature map contains feature information such as heart rate, blood pressure, movement and sleep at different time scales, and the inter-modal dependency graph contains the correlation weight information between the features of each modality, after splicing the two, the feature weight is adjusted through the gate fusion unit. If the blood pressure is more critical to the health status at the current stage, the feature weight related to the blood pressure will be appropriately increased during fusion, for example, from the original 0.4 to 0.5. Then the adjusted features are input into the bidirectional gated recurrent network for context-aware feature enhancement. The bidirectional gated recurrent network can better understand the health status features by using the context information. For example, by combining the previous health status of the middle-aged man and the current state change, the health status features are enhanced to obtain the enhanced fusion features.

[0060] At step S134, the enhanced fusion features are subjected to hierarchical pooling operation to extract a global health status summary vector and a local abnormal pattern feature vector respectively, and the global health status summary vector and the local abnormal pattern feature vector are integrated by a fully connected layer to generate the multi-granularity health status representation vector containing multi-level health information.

[0061] For example, the global health status summary vector is first extracted, which can reflect the overall health status of the middle-aged man, for example, the overall health risk level is at a medium level, which may be because he has mild hypertension but the risk has not further deteriorated through certain exercise and diet control. At the same time, the local abnormal pattern feature vector is extracted, which can highlight the local abnormal situation that he occasionally feels dizzy. Finally, the global health status summary vector and the local abnormal pattern feature vector are integrated by a fully connected layer to generate a multi-granularity health status representation vector containing multi-level health information, which can comprehensively describe the health status of the middle-aged man, including the overall health risk level and the local abnormal feature information.

[0062] In one 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:

[0063] At step S141, the cardiovascular disease risk probability, metabolic syndrome risk index and mental stress assessment score are calculated in parallel in the multi-head risk prediction module, wherein each risk prediction module uses an independent feature transformation layer and a risk classifier, and shares the bottom features of the multi-granularity health status representation vector.

[0064] In this embodiment, the multi-granularity health status representation vector of the middle-aged man 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, metabolic syndrome risk index and mental stress assessment score are calculated in parallel.

[0065] For the calculation of cardiovascular disease risk probability, since the middle-aged man has mild hypertension with blood pressure around 140 / 90 mmHg, which is an important risk factor for cardiovascular disease. At the same time, his weight is 75 kg, body fat rate is 25%, and age is middle-aged, these factors are combined to calculate through specific algorithm and model. The feature transformation layer in the model will transform the relevant features in the multi-granularity health state representation vector, such as normalizing, weighting, etc. for blood pressure, weight, age and other features, and then the risk classifier calculates the cardiovascular disease risk probability according to the transformed features, assuming the calculation result is 30%.

[0066] In the calculation of metabolic syndrome risk index, his weight, body fat rate, and exercise and diet are considered. His weight is on the high side, and his body fat rate is also at a high level, although he has certain exercise habits, but the intensity and time of exercise may not be enough to completely reduce the risk of metabolic syndrome. These relevant features are processed by independent feature transformation layers, such as re-encoding weight, body fat rate, and exercise time, and then calculated by the corresponding risk classifier, and the metabolic syndrome risk index may be 40%.

[0067] The calculation of mental stress assessment score is based on the relevant information in the behavior activity record data and medical history text data. From the behavior activity record data, his sleep quality index shows that he sleeps about 2-3 hours a night, and the shallow sleep time is longer, which may indicate that he has certain mental stress. At the same time, the occasional dizziness symptoms in the medical history may also be related to mental stress. In this multi-head risk prediction module, the feature transformation layer transforms the sleep quality, dizziness symptoms and other relevant features, and then the risk classifier calculates the mental stress assessment score, which may be 20%. Each risk prediction module uses independent feature transformation layer and risk classifier, but shares the bottom layer features of multi-granularity health state representation vector, which contains the basic information of the overall health status of the middle-aged man, such as physiological parameters, behavior activity features, etc.

[0068] Step S142, according to the weighted combination result of the cardiovascular disease risk probability and the metabolic syndrome risk index, determine the severity classification of the disease risk level label, and dynamically calibrate the classification result combined with the mental stress assessment score.

[0069] 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% x 0.6 + 40% x 0.4) = 34%. Then, the grading result is dynamically calibrated in combination with the mental stress assessment score. Since the mental stress assessment score is 20%, although not very high, it will also have a certain impact on the overall health status. After the specific calibration algorithm adjusts the previous 34% result, the disease risk level of the middle-aged man is finally determined as a medium risk. This calibration process considers the potential impact of mental stress on overall health risk, making the assessment of disease risk level more comprehensive and accurate.

[0070] Step S143, in the intervention strategy generation module, based on the calibrated disease risk level, retrieve the matched candidate intervention items from the pre-set intervention measure knowledge base, and calculate the execution feasibility score of each intervention item according to the historical behavior activity record data of the target user.

[0071] Because the middle-aged man has mild hypertension and the disease risk level is medium, the candidate intervention items include dietary adjustment (such as reducing salt intake), exercise prescription (increasing aerobic exercise time and intensity), and medical examination plan (regularly measuring blood pressure, conducting blood lipid examination, etc.).

[0072] According to the historical behavior activity record data of the middle-aged man, the execution feasibility score of each intervention item is calculated. He has certain exercise habits, such as morning walking, so the execution feasibility score of the exercise prescription is relatively high. For dietary adjustment, considering his eating habits and social activities, if he often goes out to eat, the execution feasibility score of reducing salt intake may be relatively low. For the medical examination plan, since he has a record of regularly checking blood pressure according to the doctor's advice, the execution feasibility score of this intervention item is high.

[0073] Step S144, using reinforcement learning strategy to prioritize the candidate intervention items, generate the intervention measure priority list containing dietary adjustment suggestion, exercise prescription and medical examination plan, and derive the dynamic monitoring period configuration parameters based on the risk probability change rate.

[0074] In detail, the reinforcement learning strategy takes into account the feasibility of each intervention item, the potential impact on reducing the disease risk, and the long-term health benefits, etc. For example, exercise prescription is ranked higher in the priority list because of its high feasibility and positive impact on blood pressure control and overall health status. Dietary adjustment, although slightly lower in feasibility, is also very important for controlling hypertension and is ranked after exercise prescription. The medical examination plan is ranked third according to its necessity and previous execution. Finally, an intervention priority list containing dietary adjustment suggestions (such as no more than 6 grams of salt intake per day), exercise prescriptions (increasing 30 minutes of brisk walking per week), and medical examination plans (comprehensive physical examination every three months) is generated. At the same time, the dynamic monitoring cycle configuration parameters are derived based on the risk probability change rate. Since the middle-aged man's risk is at a moderate level and relatively stable, the risk probability change rate is slow, and the dynamic monitoring cycle is configured to collect his physiological monitoring data and behavior activity record data every two weeks, and update the medical history data every three months.

[0075] In one possible implementation, step S150 specifically includes:

[0076] Step S151, when the severity classification of the disease risk level label is monitored to exceed the 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 motion posture sensor activations in the behavior activity record data is increased according to the set frequency increase strategy.

[0077] When the severity classification of the middle-aged man's disease risk level label is monitored to exceed the preset threshold, assuming that the preset threshold is 50%. If in a certain monitoring cycle, due to his blood pressure suddenly rising to 160 / 100 mmHg or new symptoms such as palpitations, etc., the disease risk level exceeds 50%. At this time, the sampling interval of the physiological monitoring data is shortened to a preset proportion of the original sampling interval, and the original comprehensive physiological data is collected every two weeks, and the sampling interval is shortened to half of the original, i.e. collected every week. At the same time, the number of motion posture sensor activations in the behavior activity record data is increased according to the set frequency increase strategy, for example, from recording the exercise state every day to recording it every half day. In this way, his exercise and health status changes can be obtained more timely, so as to more closely monitor his health status.

[0078] Step S152, dynamically updating the standardized health feature sequence according to the real-time collected multi-modal data stream, using a sliding window mechanism to perform incremental feature alignment processing on the newly added data segment, and preserving the context association of the historical feature sequence.

[0079] When new data arrives, such as newly collected data for a week, a sliding window mechanism is used to perform incremental feature alignment processing on the new data segment. The newly collected data includes new physiological monitoring data (such as heart rate, blood pressure, etc.), behavior activity record data (such as exercise status, sleep quality, etc.), and possibly updated medical history data (if there are new diagnoses or symptom descriptions). 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 the continuity of the data. For behavior activity record data, the new exercise status data needs to be arranged according to the same time reference as the previous exercise status data. In this way, on the basis of maintaining the contextual association of historical feature sequences, 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.

[0080] Step S153, input the updated standardized health feature sequence into the spatio-temporal feature extraction network for online feature abstraction, and trigger the real-time inference process of the health risk prediction model to generate an updated personalized health management plan.

[0081] The spatio-temporal feature extraction network processes the new standardized health feature sequence and recalculates the multi-granularity health state representation vector. For example, the increase in blood pressure in the new physiological monitoring data may be reflected in the multi-granularity health state representation vector as a change in the feature value related to the risk of cardiovascular disease. Then the new multi-granularity health state representation vector is input into the health risk prediction model for calculation again. The multi-head risk prediction module recalculates the cardiovascular disease risk probability, metabolic syndrome risk index, and mental stress assessment score. Due to the increase in blood pressure, the cardiovascular disease risk probability may rise to 40%. According to the new risk probability and other related indicators, the disease risk level is re-determined, such as being changed to high risk. The intervention strategy generation module re-searches and adjusts the priority list of intervention measures according to the new disease risk level, for example, increases the frequency of medical examinations, adjusts the intensity of exercise prescriptions, etc., and re-derives the dynamic monitoring period configuration parameters, such as shortening the overall collection of physiological monitoring data and behavior activity record data to every week, and updating the medical history data every two months.

[0082] Step S154, when the disease risk level label of a continuous preset number of monitoring periods remains stable, gradually extend the data collection interval until the baseline frequency is restored, and reduce the intensity of the medical examination plan in the intervention measure priority list.

[0083] When the disease risk level label of a continuous preset number of monitoring periods (assuming three periods) remains stable, for example, a middle-aged man's blood pressure stabilizes at about 130 / 80 mmHg after a period of adjustment and treatment, and the disease risk level remains at a moderate risk level. At this time, gradually extend the data collection interval until the baseline frequency is restored. Gradually restore the physiological data collection from once a week to once every two weeks, while reducing the intensity of medical examination plans in the intervention measure priority list, such as from a comprehensive physical examination every two months to once every six months. In this way, under the premise of effectively monitoring the health status of the middle-aged man, unnecessary data collection and medical examinations are reduced, resource utilization efficiency is improved, and a closed-loop health management data flow that dynamically adjusts according to the health status is formed.

[0084] In a possible implementation, the method further includes:

[0085] Step S210, after generating the personalized health management scheme, extracting a key feature activation path from the spatio-temporal feature extraction network, and constructing a feature contribution degree heat map to visualize the influence degree of different health data on the risk prediction result.

[0086] After generating the personalized health management scheme for the middle-aged man, a key feature activation path can also be extracted from the spatio-temporal feature extraction network, and a feature contribution degree heat map can be constructed to visualize the influence degree of different health data on the risk prediction result. Taking the middle-aged man as an example, when the spatio-temporal feature extraction network processes his health data, the blood pressure data in the physiological monitoring data can be a key feature for the prediction of the cardiovascular disease risk probability. When constructing the feature contribution degree heat map, the area corresponding to the blood pressure data on the heat map can show a relatively high contribution degree value, indicating that the blood pressure has an important influence on the cardiovascular disease risk prediction result. Similarly, the exercise time and sleep quality indicators in the behavior activity record data also contribute to the risk prediction to a certain extent. If his exercise time is short and the sleep quality is poor, the areas corresponding to these two features on the feature contribution degree heat map will show the corresponding contribution degree values, which intuitively show the high and low degrees of influence of these health data on the risk prediction result through different colors or value sizes.

[0087] Step S220, generating an execution effect simulation curve for each intervention item in the intervention measure priority list to show the change trend of the expected risk probability under different execution intensities.

[0088] For example, for the dietary adjustment recommendation intervention item, assuming that a middle-aged man currently consumes 8 grams of salt per day, and the salt intake is gradually reduced to 6 grams, 5 grams, 4 grams, and other different levels according to different salt intake adjustment execution intensities, the expected risk probability will change accordingly. When the salt intake is reduced to 6 grams, the cardiovascular disease risk probability may be reduced from 30% to 28% due to the positive effect on blood pressure; when the salt intake is further reduced to 5 grams, the risk probability may be reduced to 26%. Plotting the expected risk probability change data at different execution intensities into an execution effect simulation curve can clearly show the potential effect of the dietary adjustment intervention item on reducing risk at different execution intensities. For the exercise prescription intervention item, such as gradually increasing the brisk walking exercise time from 30 minutes to 40 minutes, 50 minutes, and other time periods per week, as the exercise time increases, the expected risk probability will also decrease accordingly, and the corresponding execution effect simulation curve is also plotted. For the medical examination plan intervention item, although it is mainly used for monitoring health status, different examination frequencies may also affect early detection of diseases and intervention timing, thereby indirectly affecting risk probability, and the corresponding execution effect simulation curve is also generated.

[0089] Step S230, integrating the feature contribution degree heat map and the execution effect simulation curve into an interactive explanation report, and converting the visualization results into user-oriented popularized health recommendation text through a natural language generation algorithm.

[0090] For example, after generating a personalized health management plan for a middle-aged man, the feature contribution degree heat map and the execution effect simulation curve can be integrated into an interactive explanation report. The interactive explanation report not only contains intuitive visual graphics, but also converts these graphic information into user-friendly popularized health recommendation text through a natural language generation algorithm.

[0091] For example, in the feature contribution degree heat map, if the blood pressure data has the greatest impact on the cardiovascular disease risk prediction result, it can be clearly pointed out in the popularized health recommendation text: “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 popularized health recommendation text will further state: “By gradually reducing daily salt intake, for example, from the current 8 grams to 6 grams, it is expected to effectively reduce your cardiovascular disease risk probability.”

[0092] In addition, the interactive interpretation report also allows users to interact with the graph, such as viewing the trend of the expected risk probability under different execution intensities by clicking or swiping, while the popularized health advice text can be updated in real time according to the user's interaction, providing personalized health guidance and advice. In this way, users not only can intuitively understand how their health data affects disease risk, but also can take corresponding health management measures according to the provided popularized health advice text, improving the effectiveness and pertinence of health management.

[0093] In a possible implementation, the method further comprises:

[0094] Step S310, embedding a feedback interface in the interactive interpretation report to receive user subjective execution difficulty score and physiological feeling description text of the health management scheme through the feedback interface.

[0095] Step S320, converting the user feedback data into a feedback feature vector, and inputting the feedback feature vector and the current health state representation vector into a feedback adaptation network to generate a management scheme adjustment coefficient.

[0096] Step S330, dynamically modifying the item order and execution intensity parameters in the intervention measure priority list according to the adjustment coefficient, and recalculating the dynamic monitoring period configuration parameters.

[0097] For example, in the interactive interpretation report, a feedback interface is embedded, through which a middle-aged man can feed back the health management scheme, including subjective execution difficulty score and physiological feeling description text. For example, the middle-aged man may feel that it is somewhat difficult to increase 30 minutes of brisk walking exercise per week according to the current exercise prescription, and he gives a high difficulty score in the subjective execution difficulty score, and mentions in the physiological feeling description text that "after increasing exercise, he feels very tired and difficult to persist". Convert the user feedback data into a feedback feature vector, which involves quantitative processing 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 such as "tired" and "difficult to persist" is extracted through natural language processing technology, and is converted into quantifiable feature values, and then is spliced with the current health state representation vector.

[0098] The management scheme adjustment coefficient reflects the degree of adjustment of the health management scheme according to user feedback. According to the management scheme adjustment coefficient, the order of items in the intervention measure priority list and the execution intensity parameter are dynamically modified, and the dynamic monitoring period configuration parameter is recalculated. For example, if a middle-aged man feedbacks that the exercise prescription is difficult to perform and the body is tired, the adjustment coefficient may indicate that the execution intensity of the exercise prescription should be reduced. Originally, the walking exercise time is increased by 30 minutes per week, and according to the adjustment coefficient, it may be modified to increase the walking exercise time by 15 minutes per week. At the same time, due to the reduction of exercise intensity, the priority of dietary adjustment suggestions can be adjusted accordingly, and more attention can be paid to dietary adjustment to balance the overall health management effect. For the dynamic monitoring period configuration parameter, due to the reduction of exercise intensity, it may have an impact on the improvement speed of the health status, so after recalculation, the comprehensive collection of physiological monitoring data and behavior activity record data originally performed once every two weeks can be adjusted to once every three weeks to adapt to the new health management strategy. In this way, the health management scheme is continuously optimized through user feedback to better meet the actual situation and needs of middle-aged men, and the effectiveness and executability of the health management scheme are improved.

[0099] In a possible implementation, the training process of the spatio-temporal feature extraction network includes:

[0100] Step S101, a training data set containing target health data samples and their corresponding disease diagnosis labels is constructed, wherein each target health data sample contains at least a standardized health feature sequence of a set time period.

[0101] In detail, for a target user group such as middle-aged men, each target health data sample contains at least a standardized health feature sequence of a set time period. For example, samples are collected from the health data of numerous middle-aged men, wherein one sample may contain a standardized health feature sequence of one month, which covers physiological monitoring data (such as heart rate, blood pressure, etc.), behavior activity record data (such as exercise state, 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 or not suffering from cardiovascular disease, diabetes, etc.

[0102] Step S102, a multi-task loss function is constructed, which contains a time series reconstruction loss, a cross-modal consistency loss, and a 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 alignment of the distribution of different modal features in the hidden space, and the disease classification loss supervises the accuracy of risk prediction.

[0103] Taking the heart rate data of middle-aged men as an example, after the spatio-temporal 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 according to these high-level features. If the original heart rate data is 70 times per minute at a certain time, the value reconstructed by the spatio-temporal feature extraction network should be as close as possible to this value. The cross-modal consistency loss constrains the alignment of the distribution of features in the hidden space of different modalities. For example, the features of blood pressure in physiological monitoring data and exercise energy consumption in behavior activity record data of different modalities should maintain consistent distribution as much as possible after being processed and converted to the hidden space by the network, so that the network can correctly understand and associate the relationship between different modalities when processing multi-modal data. The 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 actual disease is cardiovascular disease but the network predicts that it is not, the disease classification loss will be large.

[0104] In step S103, a pre-training model of the spatio-temporal feature extraction network is generated by gradually increasing the time span and modality complexity of the training data set using a curriculum learning strategy. In the initial training stage, short-term data of a single modality is used, and in the final stage, long-term continuous data of all modalities is used.

[0105] In one possible implementation, step S103 specifically includes:

[0106] In step S1031, an initial training stage data set corresponding to the training data set is obtained, and the initial training stage data set only contains short-term time sequence segments of a single health monitoring modality. The continuous time length of the short-term time sequence 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, behavior activity record data, or medical history text data.

[0107] For example, physiological monitoring data can be selected as a single health monitoring modality, and the continuous time length of the short-term time sequence segments does not exceed a preset initial time window threshold. Assuming that the preset initial time window threshold is one week, the initial training stage data set is composed of time sequence segments containing only physiological monitoring data (such as heart rate, blood pressure, etc.) within one week, which are selected from the health data of numerous middle-aged men.

[0108] In step S1032, a first training batch is generated based on the initial training stage data set, and the first training batch is input into the spatio-temporal feature extraction network for feature extraction capability pre-training, and an initial network weight parameter set is output. During the pre-training process, the parameters of the network layers other than the input layer are frozen.

[0109] In this pre-training process, the network layer parameters other than the input layer are frozen, the main purpose is to let the spatio-temporal feature extraction network learn the basic feature representation of single modal data first. For example, the spatio-temporal feature extraction network first learns the basic fluctuation pattern and feature relationship of heart rate and blood pressure in physiological monitoring data, and outputs an initial network weight parameter set.

[0110] Step S1033, the modal types of the initial training phase data set are expanded to two target modalities, and the time length of the short-term time sequence segment is kept unchanged, and a second training batch is generated, wherein the combination of the two target modalities is selected from physiological monitoring data and behavior activity record data, physiological monitoring data and medical history text data, or behavior activity record data and medical history text data.

[0111] Step S1034, based on the initial network weight parameter set, the spatio-temporal feature extraction network is parameterized, the second training batch is used for cross-modal correlation training, the network layer parameters related to the cross-modal attention mechanism in the spatio-temporal feature extraction network are updated, and an intermediate network weight parameter set is generated.

[0112] For example, the physiological monitoring data and the behavior activity record data are selected as the two target modalities, and the physiological monitoring data (such as heart rate, blood pressure, etc.) and the behavior activity record data (such as the number of steps, sleep quality, etc.) within a week are combined to form a second training batch. Based on the initial network weight parameter set obtained before, the spatio-temporal feature extraction network is parameterized, and the second training batch is used for cross-modal correlation training. In this process, the focus is to let the network learn the correlation between the two different modalities, update the network layer parameters related to the cross-modal attention mechanism in the spatio-temporal feature extraction network, for example, learn the potential correlation between heart rate and the number of steps, and the relationship between blood pressure and sleep quality, etc., and generate an intermediate network weight parameter set.

[0113] Step S1035, the time length of the second training batch is expanded to a medium-term time window threshold to generate a third training batch, wherein the medium-term time window threshold is a preset multiple of the initial time window threshold, and the newly added time period of cross-modal data is time-aligned and missing value interpolated during the expansion process.

[0114] Step S1036, based on the intermediate network weight parameter set, the spatio-temporal feature extraction network is parameterized, the third training batch is used for long-time sequence dependent training, the network layer parameters related to the time convolution operation in the spatio-temporal feature extraction network are updated, and a reinforced network weight parameter set is generated.

[0115] The time length of the second training batch is extended to a medium-term time window threshold, assuming that the medium-term time window threshold is three times the initial time window threshold (one week), i.e. three weeks. During the extension, the newly added time period of cross-modal data is time-aligned and missing values are imputed. For example, for the newly added two weeks of physiological monitoring data and behavior activity record data, ensure that they are correctly aligned in time with the previous one week of data, and if there are missing values (such as missing steps of a certain day), use appropriate imputation methods (such as mean imputation) to supplement. Based on the intermediate network weight parameter set, the spatio-temporal feature extraction network is loaded with parameters, and the third training batch containing three weeks of data is used for long-term sequential dependence training. During this training process, the network learns the feature change pattern under long-term sequential dependence, updates the network layer parameters related to time convolution operation in the spatio-temporal feature extraction network, such as learning the change rule of heart rate over time and the long-term trend of steps in three weeks, etc., to generate a reinforced network weight parameter set.

[0116] Step S1037, the modality type of the third training batch is expanded to the full modality type, and its time length is simultaneously expanded to the long-term time window threshold to generate a fourth training batch, wherein the full modality type includes physiological monitoring data, behavior activity record data and medical history text data, and the long-term time window threshold covers at least a complete health monitoring period of a preset number of consecutive days.

[0117] Step S1038, based on the reinforced network weight parameter set, the spatio-temporal feature extraction network is parameterized, the fourth training batch is used for multi-modal long-term sequential joint training, and all trainable parameters in the spatio-temporal feature extraction network are updated synchronously to generate a final network weight parameter set.

[0118] Step S1039, based on the final network weight parameter set, the architecture parameters of the spatio-temporal feature extraction network are updated to generate a pre-trained model adapted to multi-modal long-term sequential health data analysis.

[0119] Assume that the long-term time window threshold covers at least a complete health monitoring period of a preset number of consecutive days (such as 30 consecutive days). The full-modal type includes physiological monitoring data (such as heart rate, blood pressure, body weight, etc.), behavioral activity record data (such as exercise state, sleep quality, etc.), and medical history text data (such as past disease diagnosis, symptom description, etc.). Based on the set of reinforcement network weight parameters, the spatio-temporal feature extraction network is initialized. The fourth training batch containing 30 days of full-modal data is used for multi-modal long-time sequence joint training. In this process, the network learns the complex relationship between multi-modal data and the feature change rule under long-time sequence, and synchronously updates all trainable parameters in the spatio-temporal feature extraction network to generate the final set of network weight parameters. Finally, based on the final set of network weight parameters, the architecture parameters of the spatio-temporal feature extraction network are updated to generate a pre-trained model adapted to multi-modal long-time sequence health data analysis. The pre-trained model can effectively process multi-modal long-time sequence health data of a target user such as a middle-aged man.

[0120] In step S104, the health population data is introduced as a negative sample through an adversarial training mechanism to enhance the sensitivity of the pre-trained model to abnormal health states, and the spatio-temporal feature extraction network is generated.

[0121] For example, in a possible implementation, step S104 specifically includes:

[0122] In step S1041, an initial set of spatio-temporal feature extraction network parameters trained by a curriculum learning strategy is obtained, and an adversarial training data set containing a health population data set and an abnormal health state sample set is obtained. The samples in the health 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.

[0123] First, the initial spatio-temporal feature extraction network parameter set after course learning strategy training is obtained, which is the pre-training result obtained by gradually increasing the time span and modality complexity of the training data set. At the same time, an adversarial training data set containing a healthy population data set and an abnormal health state sample set is prepared. Taking middle-aged men as an example, the middle-aged male samples in the healthy population data set have normal health indicators, such as heart rate of 60-70 times / minute, blood pressure of about 120 / 80 mmHg, normal exercise state (such as moderate exercise every day, and weekly exercise time reaching a certain standard) and no major disease medical history, etc. These samples are labeled as negative sample labels, indicating a healthy state. The middle-aged male samples in the abnormal health state sample set may have cardiovascular diseases, their heart rate may often exceed 80 times / minute, their blood pressure may be above 140 / 90 mmHg, their exercise state may be unstable due to disease (such as decreased exercise tolerance, easy fatigue), and they have a diagnosis record of cardiovascular disease in their medical history. These samples are labeled as positive sample labels, indicating a diseased state.

[0124] Step S1042, constructing an adversarial sample generator network, inputting the healthy population data set into the adversarial sample generator network, and generating synthetic health data flow with similar statistical distribution to abnormal health state samples but labeled as negative sample labels.

[0125] The adversarial sample generator network generates synthetic health data flow with similar statistical distribution to abnormal health state samples but labeled as negative sample labels by learning the statistical distribution characteristics of abnormal health state samples. For example, for heart rate data in the healthy population data, although the overall is in the normal range, the adversarial sample generator network will generate heart rate data in special cases according to the fluctuation pattern of the heart rate of the abnormal health state sample, the association with other physiological indicators, etc. Characteristics, such as heart rate fluctuation data under the influence of simulating some potential health risk factors, but these data are still within the normal range, thereby forming synthetic health data flow.

[0126] Step S1043, mixing the synthetic health data flow and the real abnormal health state sample set according to a predetermined proportion to generate an adversarial training batch data, and inputting the adversarial training batch data into the initial spatio-temporal feature extraction network for feature extraction to obtain an initial health state feature set.

[0127] Assuming 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 input the adversarial training batch data into the initial spatio-temporal feature extraction network for feature extraction to obtain an initial health state feature set. In this process, the spatio-temporal feature extraction network processes the mixed sample data, extracts the feature representation of each sample according to its existing network structure and parameters, and these features cover the comprehensive features of physiological monitoring data, behavior activity record data, medical history text data and other multi-modal information, forming an initial health state feature set.

[0128] Step S1044, input the initial health state feature set into the adversarial discriminator network, calculate the discrimination loss value of the adversarial discriminator network for positive and negative samples, 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.

[0129] If the adversarial discriminator network incorrectly determines the sample in the synthetic health data stream as a positive sample (i.e., determines it as an abnormal health state), or determines the real abnormal health state sample as a negative sample (i.e., determines it as a healthy state), a larger 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 updated. For example, if a sample in a certain synthetic health data stream is misjudged as an abnormal health state, it indicates that the sample generated by the adversarial sample generator network deviates from the real health population data distribution, and the generation parameters of the sample need to be adjusted to make the sample generated by the adversarial sample generator network more consistent with the characteristics of the health population data; at the same time, the discrimination parameters of the adversarial discriminator network also need to be adjusted to improve its discrimination ability for positive and negative samples.

[0130] Step S1045, the updated adversarial sample generator network regenerates a new synthetic health data stream, and iteratively performs the mixing of the adversarial training batch data, the feature extraction and the discrimination loss calculation process until the adversarial discriminator network cannot distinguish the real abnormal sample from the synthetic health data at a preset accuracy threshold.

[0131] Assuming that the preset accuracy threshold is 90%, in the iteration process, as the adversarial sample generator network is continuously optimized, the synthetic health data stream generated by the adversarial sample generator network is closer and closer to the real health population data distribution, and the discrimination difficulty of the adversarial discriminator network for positive and negative samples gradually increases. When the discrimination accuracy of the adversarial discriminator network for the sample is close to 90%, it means that the synthetic health data stream and the real abnormal health state sample are very similar in features and are difficult to accurately distinguish.

[0132] Step S1046, freeze the parameters of the adversarial sample generator network, input the healthy population data set and the abnormal health state sample set into the adversarial sample generator network together, and generate an adversarial enhanced feature set, wherein the adversarial enhanced feature set contains negative sample feature disturbance items and positive sample feature strengthening items.

[0133] In this process, since the parameters of the adversarial sample generator network have been fixed, it will perform specific processing on the input healthy population data and abnormal health state samples, generating an adversarial enhanced feature set containing negative sample feature disturbance items and positive sample feature strengthening items. For healthy population data, for example, normal blood pressure data may be added with some small negative sample feature disturbance items, so that the representation of the healthy population data in the feature space is more diversified, which helps the adversarial sample generator network to better distinguish between the healthy population and the abnormal health state; for abnormal health state samples, such as high blood pressure data of middle-aged men with cardiovascular diseases, positive sample feature strengthening items will be added to highlight the differences between them and the healthy population data in the feature space, for example, to strengthen the abnormal correlation between blood pressure and other physiological indicators.

[0134] Step S1047, feature-level splicing the adversarial enhanced feature set and the initial health state feature set to generate an adversarial training feature matrix, and inputting the adversarial training feature matrix into the spatio-temporal feature extraction network for feature reconstruction training.

[0135] The adversarial training feature matrix contains the comprehensive feature information of the specially processed healthy population and abnormal health state samples. Then the adversarial training feature matrix is input into the spatio-temporal feature extraction network for feature reconstruction training. The spatio-temporal feature extraction network attempts to reconstruct an output similar to the original adversarial training batch data according to the adversarial training feature matrix. For example, for the health data of middle-aged men, if the original adversarial training batch data contains heart rate, blood pressure, exercise state and medical history information at a specific time point, the spatio-temporal feature extraction network will try to reconstruct these information as accurately as possible.

[0136] Step S1048, calculate the distribution consistency loss between the reconstruction output of the spatio-temporal feature extraction network and the original adversarial training batch data, and combine the feature confusion loss provided by the adversarial discriminator network to generate an adversarial training total loss value.

[0137] Step S1049, according to the adversarial training total loss value, gradient optimization is performed on the parameters of the spatio-temporal feature extraction network, and the convolution kernel weight and attention mechanism parameters related to cross-modal feature extraction in the spatio-temporal feature extraction network are updated.

[0138] If the distribution of the reconstructed data such as heart rate and blood pressure by the spatio-temporal feature extraction network is significantly different from the original data, a larger distribution consistency loss will be generated. At the same time, if the reconstructed features are highly confused by the adversarial discriminator network, the feature confusion loss will also increase. The total adversarial training loss value is obtained by adding the two, and the parameters of the spatio-temporal feature extraction network are optimized according to the total adversarial training loss value, and the convolution kernel weights and attention mechanism parameters related to cross-modal feature extraction in the spatio-temporal feature extraction network are mainly updated. For example, adjust the convolution kernel weights to better capture the relationship between different modal data, and optimize the attention mechanism parameters to improve the attention degree to key features.

[0139] Step S10410, dynamically adjust the proportion of synthetic healthy data stream and real abnormal samples in the adversarial training batch data, so that the proportion of the synthetic healthy data stream decreases with the increase of the training round in the training process, until it is completely replaced by the real abnormal health state sample set.

[0140] At the beginning, the proportion of synthetic healthy data stream and real abnormal samples may be 1:1, and with the increase of the training round, it is gradually adjusted to 1:2, 1:3, etc., until the last training is completely performed using the real abnormal health state sample set. In this process, the operations of adversarial sample generation, feature enhancement, discriminant loss calculation and network parameter updating are repeatedly performed until the class distance in the hidden space between the feature extraction results of the healthy population data set and the feature extraction results of the abnormal health state sample set by the spatio-temporal feature extraction network exceeds the preset threshold. Assuming that the preset threshold is 0.8, it means that the spatio-temporal feature extraction network can well distinguish the feature representation of healthy population and abnormal health state in the hidden space, that is, the class distance between the two is large enough.

[0141] Step S10411, repeatedly perform the operations of adversarial sample generation, feature enhancement, discriminant loss calculation and network parameter updating until the class distance in the hidden space between the feature extraction results of the healthy population data set and the feature extraction results of the abnormal health state sample set by the spatio-temporal feature extraction network exceeds the preset threshold.

[0142] Step S10412, decouple the final optimized spatio-temporal feature extraction network parameters and the parameters of the adversarial discriminator network, and only keep the spatio-temporal feature extraction network parameters as the enhanced model for abnormal health state recognition.

[0143] Step S10413, verify the classification accuracy of the enhanced model on the independent test set for healthy population and abnormal health state, and when the classification accuracy does not reach the preset performance threshold, readjust the perturbation intensity parameter of the adversarial sample generator network and iteratively perform the adversarial training process.

[0144] The independent test set includes health data samples of middle-aged men who did not participate in the training, both healthy samples and abnormal health state samples. If the classification accuracy does not reach the preset performance threshold (for example, 95% is preset), the perturbation strength parameter of the adversarial sample generator network is adjusted and the adversarial training process is iteratively performed. For example, if the classification accuracy is 90%, the perturbation strength of the adversarial sample generator network is appropriately increased, and the above-mentioned adversarial training process is performed again until the classification accuracy reaches or exceeds 95%.

[0145] In step S10414, the enhanced model that passes the verification is parameter fused with the pre-trained model trained by the curriculum learning strategy to generate a spatio-temporal feature extraction network with cross-modal abnormal sensitivity.

[0146] In this embodiment, the final spatio-temporal feature extraction network has both the processing capability for multi-modal long-time series health data obtained through the curriculum learning strategy and the recognition sensitivity for abnormal health states enhanced through the adversarial training mechanism, so as to more accurately process the health data of the target user such as middle-aged men.

[0147] In a possible implementation, the method further includes:

[0148] In step S410, after the training is completed, the spatio-temporal feature extraction network is subjected to channel pruning and quantization-aware training to remove redundant feature channels and convert floating-point weights to fixed-point number representation.

[0149] For example, in the spatio-temporal feature extraction network, there may be channels in some convolutional layers that contribute less to the health data feature representation. By analyzing the importance of each channel in processing the health data of middle-aged men, for example, observing the degree of contribution to the extraction of key features such as heart rate and blood pressure, those channels with less contribution are removed. Quantization-aware training is to convert floating-point weights to fixed-point number representation. In processing the health data of middle-aged men, the weights in the network may be floating-point data, and quantization-aware training converts these weights to fixed-point numbers, reduces storage space and computational complexity without losing too much accuracy, and improves the running efficiency of the network.

[0150] In step S420, a lightweight feature caching mechanism is constructed to store the preprocessing results of the standardized health feature sequence and the low-dimensional health state representation vector in the local memory of the edge computing device.

[0151] For the health data of middle-aged men, the standardized health feature sequence after preprocessing, such as the results after the aforementioned noise filtering, feature alignment and other operations, and the low-dimensional health state representation vector obtained through the spatio-temporal feature extraction network, are all 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 scheme generation is needed, these data can be directly obtained from the local memory, reducing data transmission and recalculation time and improving system response speed.

[0152] Step S430, constructing a multi-thread reasoning engine, decomposing the calculation tasks of the health risk prediction model into sub-tasks executed in parallel, realizing real-time generation and update of health management schemes.

[0153] Taking middle-aged men as an example, when the health risk prediction model calculates the cardiovascular disease risk probability, metabolic syndrome risk index and mental stress assessment score, etc., these tasks can be decomposed into multiple sub-tasks. For example, the sub-task of calculating the cardiovascular disease risk probability can be performed simultaneously with the sub-task of calculating the metabolic syndrome risk index. The multi-thread reasoning engine allocates computing resources reasonably, so that these sub-tasks are executed in parallel, thereby speeding up the calculation of the health risk prediction model, and being able to generate or update the health management scheme in time according to the latest health data of the middle-aged man, such as adjusting the intervention measure priority list or dynamically monitoring the cycle configuration parameters, etc., to realize real-time monitoring and effective management of the health status of the middle-aged man.

[0154] Figure 2 The hardware structure of the health management platform system 100 provided by the embodiments of the present application is shown. As shown in the figure, the health management platform system 100 can include a processor 110, a machine readable storage medium 120, a bus 130, and a communication unit 140. Figure 2

[0155] In a 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 the machine readable storage medium 120 via a network. For another example, the health management platform system 100 can be directly connected to the machine readable storage medium 120 to access the stored information and / or data. In some embodiments, the health management platform system 100 can be implemented on a health management platform system. For example only, the health management platform system can include a private cloud, a semantic-related cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any aggregation thereof.​

[0156] The machine readable storage medium 120 can store data and / or instructions. In some embodiments, the machine readable storage medium 120 can store data acquired from an external terminal. In some embodiments, the machine readable storage medium 120 can store data and / or instructions used by the health management platform system 100 to perform or use to complete the exemplary methods described in the present application.

[0157] In the implementation process, the one or more processors 110 execute the computer executable instructions stored in the machine readable storage medium 120, so that the processor 110 can perform the deep learning based health management data mining method of the method embodiments as described above. 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 transceiving action of the communication unit 140.

[0158] The implementation process of the processor 110 can refer to the above-mentioned various method embodiments executed by the health management platform system 100, and the implementation principle and technical effects are similar, which will not be described here again.

[0159] In addition, the embodiment of the present application also provides a readable storage medium, wherein the readable storage medium is provided with computer executable instructions, and when the processor executes the computer executable instructions, the deep learning based health management data mining method as described above is realized.

[0160] It should be noted that, in order to simplify the description of the present application and help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof. Similarly, it should be noted that, in order to simplify the description of the present application and help understand one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A deep learning-based health management data mining method, characterized by, The method comprises: acquiring a target health data set of a target user, the target health data set comprising physiological monitoring data, behavior activity record data, and medical history text data, wherein the physiological monitoring data contains multiple types of continuous physiological parameters, the behavior activity record data contains motion state and sleep quality indicators in time series, and the medical history text data contains 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 contains a synchronous timestamp, a cross-modal correlation identifier, and a normalized numerical representation; performing hierarchical feature abstraction on the standardized health feature sequence through a cascaded spatio-temporal feature extraction network to obtain a multi-granularity health state representation vector, wherein the spatio-temporal feature extraction network comprises a parallelly arranged time series convolution branch and a spatial attention branch, the time series convolution branch is used to capture long-period health trends, and the spatial attention branch is used to identify dynamic dependency relationships between cross-modal features; inputting the multi-granularity health state representation vector into a pre-trained health risk prediction model to generate an individualized health management scheme for the target user, the individualized health management scheme comprising a disease risk level label, an intervention measure priority list, and dynamic monitoring period configuration parameters; adjusting the acquisition frequency and feature alignment strategy of the target health data set in real time according to the dynamic monitoring period configuration parameters in the individualized health management scheme to form a closed-loop health management data flow; the hierarchical feature abstraction on the standardized health feature sequence through the cascaded spatio-temporal feature extraction network to obtain a multi-granularity health state representation vector specifically comprises: in the time series convolution branch, a multi-scale time series pattern extraction is performed on the standardized health feature sequence by using a dilated causal convolution layer stacking structure to obtain a multi-scale time series feature map, wherein convolution layers with different dilation rates capture short-term physiological fluctuations, medium-term behavior patterns, and long-term health trends, respectively; in the spatial attention branch, a cross-modal feature correlation matrix is constructed, the contribution degrees of different modal features to the health state are calculated through a learnable attention weight distribution mechanism, and an inter-modal dependency graph is generated; the multi-scale time series feature map output by the time series convolution branch and the inter-modal dependency graph generated by the spatial attention branch are spliced, and after adjusting the feature weights through a gate fusion unit, the enhanced fusion features are input into a bidirectional gated recurrent network for context-aware feature enhancement to obtain enhanced fusion features; perform hierarchical pooling operation on the enhanced fusion features to extract a global health state summary vector and a local abnormal pattern feature vector, respectively, and generate the multi-granularity health state representation vector containing multi-level health information through dimension reduction integration through a fully connected layer. 2.The deep learning-based health management data mining method of claim 1, wherein, the cross-modal noise filtering and feature alignment processing on the target health data set to generate a standardized health feature sequence specifically comprises: Abnormal fluctuation detection is performed on continuous physiological parameters in the physiological monitoring data, an adaptive threshold segmentation algorithm is used to identify signal collection noise segments, and data repair is performed on the signal collection noise segments through bidirectional circular interpolation to obtain repaired physiological parameters; Motion state indicators in the behavior activity record data are classified by action type, discrete action events are converted into continuous energy consumption curves based on a preset motion energy consumption mapping table, and the sleep quality indicators are time window aligned to obtain aligned behavior activity curves; Unstructured symptom descriptions in the medical history text data are subjected to entity relationship extraction, a triple graph knowledge graph of symptoms-diagnosis-treatment is constructed, and graph nodes in the triple graph knowledge graph are embedded into low-dimensional semantic vectors to obtain semantic vectorized medical knowledge; The repaired physiological parameters, aligned behavior activity curves, and semantic vectorized medical knowledge are subjected to multi-modal feature fusion, a dynamic time warping algorithm is used to eliminate time offset of cross-modal data, and the standardized health feature sequence with a unified time reference is generated. 3.The deep learning-based health management data mining method of claim 1, wherein, 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 an individualized health management scheme for the target user, specifically including: In the multi-head risk prediction module, cardiovascular disease risk probability, metabolic syndrome risk index, and mental stress assessment score are calculated in parallel, wherein each risk prediction module uses an independent feature transformation layer and a risk classifier, and shares the bottom features of the multi-granularity health state representation vector; According to the weighted combination result of the cardiovascular disease risk probability and the metabolic syndrome risk index, the severity classification of the disease risk level label is determined, and the classification result is dynamically calibrated 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 according to the historical behavior activity record data of the target user; The reinforcement learning strategy is used to prioritize the candidate intervention items, generate the intervention measure priority list containing dietary adjustment suggestions, exercise prescriptions, and medical examination plans, and derive the dynamic monitoring cycle configuration parameters based on the risk probability change rate. 4.The deep learning-based health management data mining method of claim 1, wherein, According to the dynamic monitoring cycle configuration parameters in the individualized health management scheme, 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, specifically including: When it is monitored that the severity classification 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 activation frequency of the motion posture sensor in the behavior activity record data is increased according to the set frequency increase strategy; The standardized health feature sequence is dynamically updated according to the real-time collected multi-modal data flow, a sliding window mechanism is adopted to perform incremental feature alignment processing on the newly added data segment, and the context correlation of the historical feature sequence is preserved; The updated standardized health feature sequence is input into the spatio-temporal feature extraction network for online feature abstraction, and a real-time inference process of the health risk prediction model is triggered to generate an updated personalized health management scheme; When the disease risk level labels of a continuous preset number of monitoring periods remain stable, 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. 5.The deep learning-based health management data mining method of claim 1, wherein, The method further comprises: After generating the personalized health management scheme, key feature activation paths are extracted from the spatio-temporal feature extraction network, a feature contribution degree heat map is constructed to visualize the influence degree of different health data on the risk prediction result; An execution effect simulation curve is generated for each intervention item in the intervention measure priority list to show the change trend of the expected risk probability under different execution intensities; The feature contribution degree heat map and the execution effect simulation curve are integrated into an interactive explanation report, and the visualized result is converted into user-oriented popularized health suggestion text through a natural language generation algorithm. 6.The deep learning-based health management data mining method of claim 5, wherein, The method further comprises: A feedback interface is embedded in the interactive explanation report to receive user subjective execution difficulty scores and physiological feeling description texts for the health management scheme through the feedback interface; User feedback data are converted into a feedback feature vector, which is input into a feedback adaptation network after being spliced with the current health state representation vector to generate a management scheme adjustment coefficient; The adjustment coefficient is used to dynamically modify the order of items and the execution intensity parameters in the intervention measure priority list, and the dynamic monitoring period configuration parameters are recalculated. 7.The deep learning-based health management data mining method of claim 1, wherein, The training process of the spatio-temporal feature extraction network comprises: A training data set containing target health data samples and their corresponding disease diagnosis labels is constructed, wherein each target health data sample contains a standardized health feature sequence of at least a continuous set time period; A multi-task loss function is constructed, which includes a time series reconstruction loss, a cross-modal consistency loss, and a 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 hidden space. 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, and a pre-trained model of the spatio-temporal feature extraction network is generated. In the initial training stage, short-term data of a single modality is used, and in the final stage, long-term continuous data of all modalities is used; Healthy population data are introduced as negative samples through an adversarial training mechanism to enhance the recognition sensitivity of the pre-trained model to abnormal health states, and the spatio-temporal feature extraction network is generated. 8.The deep learning-based health management data mining method of claim 7, wherein, The curriculum learning strategy gradually increases the time span and modal complexity of the training data set to generate a pre-trained model of the spatio-temporal feature extraction network, specifically comprising: obtaining an initial training phase data set corresponding to the training data set, the initial training phase data set only containing 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 behavior activity record data or the medical history text data; generating a first training batch based on the initial training phase data set, inputting the first training batch into the spatio-temporal feature extraction network for pre-training of feature extraction capability, and outputting an initial network weight parameter set, wherein during the pre-training process, the network layer parameters other than the input layer are frozen; extending the modality type of the initial training phase data set to two target modalities while keeping the time length of the short-term time series segments unchanged to generate a second training batch, wherein the combination of the two target modalities is selected from physiological monitoring data and behavior activity record data, physiological monitoring data and medical history text data, or behavior activity record data and medical history text data; based on the initial network weight parameter set, performing parameter initialization on the spatio-temporal feature extraction network, performing cross-modal correlation training using the second training batch, updating the network layer parameters related to the cross-modal attention mechanism in the spatio-temporal feature extraction network, and generating an intermediate network weight parameter set; extending the time length of the second training batch to a medium-term time window threshold to generate a third training batch, wherein the medium-term time window threshold is a preset multiple of the initial time window threshold, and during the extension process, the newly added time period of cross-modal data is time-aligned and missing value interpolated; based on the intermediate network weight parameter set, performing parameter loading on the spatio-temporal feature extraction network, performing long-time sequence dependent training using the third training batch, updating the network layer parameters related to the time convolution operation in the spatio-temporal feature extraction network, and generating a reinforced network weight parameter set; extending the modality type of the third training batch to a full modality type and synchronously extending 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, behavior activity record data and medical history text data, and the long-term time window threshold covers at least a complete health monitoring period of a preset number of consecutive days; based on the reinforced network weight parameter set, performing parameter initialization on the spatio-temporal feature extraction network, performing multi-modal long-time sequence joint training using the fourth training batch, and synchronously updating all trainable parameters in the spatio-temporal feature extraction network to generate a final network weight parameter set; based on the final network weight parameter set, updating the architecture parameters of the spatio-temporal feature extraction network to generate a pre-trained model adapted to multi-modal long-time sequence health data analysis.

9. A health management platform system, characterized by, The health management platform system comprises a processor and a memory, the memory and the processor are connected, the memory is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the memory to realize the deep learning-based health management data mining method in any one of claims 1-8.

Images