Health care data processing method and system based on intelligent perception

Abstract

The invention relates to the technical field of data processing, and discloses a health care data processing method and system based on intelligent perception. The method comprises the steps that a health state dynamic model is constructed, multi-modal physiological data and environment perception data are collected and fused to construct the model, and the model is dynamically updated; performing health risk prediction, and outputting a future health risk level by using a model based on a time sequence attention mechanism; a dynamic health intervention strategy is designed, and personalized health care suggestions are generated according to the risk levels; and performing privacy protection and model optimization, training the model by means of a federated learning framework, and ensuring data security. The system correspondingly comprises a health state dynamic modeling module, a health risk prediction module, a dynamic health intervention strategy module and a privacy protection and model optimization module. According to the invention, accurate monitoring, risk prediction and personalized intervention of the individual health state are realized, the data security is ensured, and the health care service quality is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and specifically to a health care data processing method and system based on intelligent perception. Background Art

[0002] As people's living standards improve and their health awareness increases, the demand for health care is growing. The traditional health care model mainly relies on manual experience and regular physical examinations, which makes it difficult to achieve real-time, accurate monitoring and personalized intervention of individual health status.

[0003] In terms of health data collection, the previous methods were relatively simple, usually focusing only on some indicators in physiological data, such as blood pressure, blood sugar, etc., ignoring the relationship between multimodal data. For example, relying solely on blood pressure data cannot fully understand an individual's cardiovascular health status. Information such as heart rate variability and body movement data is equally important. At the same time, the impact of environmental factors on health is often overlooked. Environmental perception data such as indoor temperature, humidity, and air quality are closely related to human health, but have not been fully utilized. This makes it impossible for the collected data to fully and accurately reflect the individual's health status, making it difficult to conduct in-depth analysis and effective use.

[0004] In terms of health risk prediction, most existing prediction models are based on static data and simple algorithms, which cannot adapt to the dynamic changes in individual health status. These models often only consider short-term data, lack effective mining of historical time series data, and cannot capture the long-term trends and potential changes of health indicators. For example, when predicting the risk of cardiovascular disease, it is not possible to make a comprehensive judgment based on the individual's long-term heart rate and blood pressure fluctuations, resulting in inaccurate prediction results and failure to provide individuals with effective health warnings in advance.

[0005] In the formulation of health intervention strategies, traditional methods lack personalization. Common health care recommendations are usually adopted without fully considering individual health risk levels, living habits, and physical condition differences. For example, giving the same exercise intensity recommendations to people with different athletic abilities and health conditions may not only fail to achieve health care effects, but may also cause harm to the body. Moreover, these intervention strategies cannot be dynamically adjusted according to the individual's real-time health data and environmental changes, and it is difficult to meet the health care needs of individuals in different scenarios.

[0006] Data security and privacy protection are also important issues facing the healthcare sector. With the collection and storage of large amounts of personal health data, the risk of data leakage increases. Once these sensitive data are leaked, it will cause serious damage to personal privacy and trigger a crisis of trust. Existing data storage and transmission methods are difficult to effectively ensure data security and privacy in the face of an increasingly complex network environment. Summary of the invention

[0007] The purpose of the present invention is to provide a health care data processing method and system based on intelligent perception to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solution: a health care data processing method based on intelligent perception, the method comprising: Constructing a dynamic health status model, including collecting multimodal physiological data and environmental perception data, and constructing a dynamically updated health status representation model by fusing the multimodal physiological data and environmental perception data; the multimodal physiological data includes heart rate, blood pressure, blood oxygen saturation and body movement data; the environmental perception data includes indoor temperature, humidity, light intensity and air quality index; Performing health risk prediction, including constructing a health prediction model based on a temporal attention mechanism, inputting historical temporal data of the multimodal physiological data and real-time environmental perception data into the health prediction model, and outputting a health risk level within a future time window; the health risk level includes a physiological abnormality probability and a potential health threat index; Designing a dynamic health intervention strategy, including generating personalized health care suggestions according to the health risk level output by the health prediction model; the dynamic health intervention strategy includes adjusting health care plan parameters, and the health care plan parameters include exercise intensity recommendations, diet plans, and rest cycles; Perform privacy protection and model optimization, including using a federated learning framework to perform distributed training on user data, updating parameters of the health prediction model, and ensuring local storage and encrypted transmission of user data.

[0009] Preferably, the construction of the health status dynamic model includes: Collect the multimodal physiological data and environmental perception data in real time through wearable devices and intelligent environmental sensors; Using a Transformer network to extract features from the multimodal physiological data to generate a time-dependent health state vector; spatially fusing the health status vector with environmental perception data to generate a comprehensive health status score through a multi-layer perceptron; The health status dynamic model is updated through a sliding window mechanism, and the window length is dynamically adjusted according to the user behavior pattern.

[0010] Preferably, the health prediction model of the temporal attention mechanism includes: Input layer: receiving historical physiological time series data and real-time environmental perception data, wherein the historical physiological time series data is divided into multiple segments according to time steps; Temporal embedding layer: extracts local features of each time segment through a one-dimensional convolutional network and generates temporal position encoding; Multi-head attention layer: assigns attention weights to the local features and temporal position encodings to capture long-term and short-term dependencies; Output layer: Map the attention weight allocation results through a fully connected network and output the probability distribution of health risk levels.

[0011] Preferably, the processing of the environmental perception data includes: Quantify the air quality index in sections and generate air quality grade labels; Identify the day and night patterns of light intensity and generate a light adaptability coefficient based on the user's daily routine data; The environmental comfort index is generated by fusing temperature, humidity and air quality level labels through fuzzy logic algorithm; The environmental comfort index is used to adjust the exercise intensity recommendation in the health care plan parameters.

[0012] Preferably, the generation of the dynamic health intervention strategy includes: Prioritize interventions based on health risk levels, with high-priority risks triggering immediate alerts and emergency recommendations; For low-priority risks, a reinforcement learning algorithm is used to generate a long-term health care plan. The reinforcement learning algorithm uses user behavior feedback as a reward signal to optimize the long-term effect of the intervention strategy; The adjustment formula of the health care plan parameters is: in, is the intervention intensity at time t, For real-time health risk level, is the environmental comfort index, is an indicator of user behavior compliance. Preferably, the implementation of the federated learning framework includes: Each terminal device locally trains a sub-model of the health prediction model and only uploads the model gradient to the central server; The central server aggregates the gradients to generate global model parameters and distributes them to each terminal; Differential privacy technology is used to add noise to uploaded gradients to ensure that user data cannot be reversed.

[0013] Preferably, the preprocessing of the multimodal physiological data includes: Wavelet transform was used to remove motion artifacts from the heart rate data; Perform Kalman filter smoothing on the blood oxygen saturation data; Extract frequency domain features from body motion data, including energy spectrum density and main frequency components; The processed multimodal data are aligned by timestamp to generate a standardized feature matrix.

[0014] Preferably, the training process of the health prediction model includes: Data collection: Collect physiological data from wearable devices of at least 1,000 users over a period of years and label health events; Feature enhancement: Expand minority health event samples through time series interpolation and adversarial generative networks; Model validation: Five-fold cross validation was used to evaluate the prediction accuracy, and the SHAP value was introduced to explain the basis for model decision making.

[0015] Preferably, the collection of user behavior feedback includes: Analyze users’ textual comments on health care suggestions through natural language processing; Capture the user's emotional state through the camera and microphone to generate an emotional polarity score; The text evaluation is integrated with the sentiment polarity score to generate a behavioral compliance index for optimizing the dynamic health intervention strategy.

[0016] Preferably, the present invention also includes a health care data processing system based on intelligent perception, the system comprising: Health status dynamic modeling module: used to collect multimodal physiological data and environmental perception data, and build a dynamically updated health status representation model by fusing the multimodal physiological data and environmental perception data. The multimodal physiological data includes heart rate, blood pressure, blood oxygen saturation and body movement data, and the environmental perception data includes indoor temperature, humidity, light intensity and air quality index; Health risk prediction module: used to build a health prediction model based on the temporal attention mechanism, input the historical temporal data of the multimodal physiological data and the real-time environmental perception data into the health prediction model, and output the health risk level in the future time window, wherein the health risk level includes the probability of physiological abnormality and the potential health threat index; Dynamic health intervention strategy module: used to generate personalized health care suggestions according to the health risk level output by the health prediction model, and the dynamic health intervention strategy includes adjusting health care plan parameters, and the health care plan parameters include exercise intensity suggestions, diet plans and rest cycles; Privacy protection and model optimization module: Use the federated learning framework to perform distributed training on user data, update the parameters of the health prediction model, and ensure the local storage and encrypted transmission of user data.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention collects multimodal physiological data and environmental perception data, and uses Transformer networks and multi-layer perceptrons to build a dynamic model of health status. The model can comprehensively and accurately reflect the health status of an individual, and greatly improves the accuracy of health monitoring compared to traditional single data monitoring methods. For example, a more comprehensive assessment of the individual's health status is made by comprehensively considering multiple aspects of information such as heart rate, blood pressure, blood oxygen saturation, body movement data, and indoor temperature, humidity, light intensity, and air quality index. At the same time, the model is dynamically updated through a sliding window mechanism, which can track changes in health status in real time and detect potential health problems in a timely manner.

[0018] The health prediction model based on the temporal attention mechanism can fully mine the historical temporal information of multimodal physiological data and real-time environmental perception data. The model effectively captures long-term and short-term dependencies through the collaborative work of the temporal embedding layer, the multi-head attention layer and the output layer, and outputs accurate health risk levels in the future time window. For example, when predicting the risk of cardiovascular disease, the long-term physiological data trend and real-time environmental factors, such as cold weather, may cause blood pressure to rise, so as to more accurately predict the probability of disease occurrence and provide strong support for early prevention and intervention. According to the health risk level output by the health prediction model, the present invention generates personalized health care suggestions, and realizes dynamic health intervention by adjusting health care plan parameters, such as exercise intensity suggestions, diet plans and rest cycles. For high-risk levels, instant alarms and emergency suggestions are triggered in time to ensure the life and health safety of individuals; for low-risk levels, a reinforcement learning algorithm is used to generate a long-term health care plan, and the intervention strategy is continuously optimized according to user behavior feedback to improve the health care effect. This personalized dynamic intervention method can better meet the health care needs of different individuals and improve the health level and quality of life of users.

[0019] The federated learning framework is used to conduct distributed training on user data. Each terminal device only uploads the model gradient, ensuring local storage and encrypted transmission of user data, effectively protecting user privacy. At the same time, the central server aggregates the gradient to generate global model parameters and distributes them to each terminal. Differential privacy technology is used to add noise to the uploaded gradient, and the model is optimized and updated under the premise of ensuring privacy. This enables the model to continuously learn new data features, improve the accuracy of prediction and intervention, and enhance the practicality and reliability of the system.

[0020] The method and system of the present invention integrate various technical advantages, from health monitoring, risk prediction to intervention strategy formulation, and then to privacy protection, forming a complete closed loop. Through comprehensive and accurate data processing and personalized services, it can provide more scientific and efficient health care solutions for health care institutions and individuals, promote the intelligent development of the health care industry, meet people's growing health needs, and have significant social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a working principle diagram of the health care data processing method based on intelligent perception according to the present invention; Figure 2 Flowcharts generated for dynamic health intervention strategies; Figure 3 Flowchart for preprocessing of multimodal physiological data; Figure 4 Diagram of the steps for training a health prediction model. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] See also Figure 1-4 The present invention provides a technical solution: a health care data processing method based on intelligent perception, the method comprising: Build a dynamic health status model: Use wearable devices and smart environmental sensors to collect multimodal physiological data (including heart rate, blood pressure, blood oxygen saturation and body movement data) and environmental perception data (including indoor temperature, humidity, light intensity and air quality index) in real time. Use the Transformer network to extract features from multimodal physiological data and generate a time-dependent health status vector. Spatially fuse the health status vector with the environmental perception data and use a multi-layer perceptron to generate a comprehensive health status score. The dynamic health status model is updated with the help of a sliding window mechanism, and the window length is dynamically adjusted based on the user's behavior pattern.

[0024] Health risk prediction: Build a health prediction model based on the temporal attention mechanism. The input layer of the model receives historical physiological time series data (divided into multiple segments according to time steps) and real-time environmental perception data; the temporal embedding layer extracts the local features of each time segment through a one-dimensional convolutional network and generates time position encoding; the multi-head attention layer allocates attention weights to the local features and time position encoding to capture long-term and short-term dependencies; the output layer maps the attention weight allocation results through a fully connected network and outputs the probability distribution of health risk levels (including the probability of physiological abnormalities and potential health threat indexes).

[0025] Design dynamic health intervention strategies: Generate personalized health care recommendations based on the health risk level output by the health prediction model. Specifically, it includes dividing intervention priorities based on health risk levels, with high-priority risks triggering immediate warnings and emergency recommendations; and using reinforcement learning algorithms to generate long-term health care plans for low-priority risks. The reinforcement learning algorithm uses user behavior feedback as a reward signal to optimize the long-term effect of the intervention strategy.

[0026] Privacy protection and model optimization: Distributed training of user data is performed using a federated learning framework. Each terminal device trains a sub-model of the health prediction model locally and only uploads the model gradient to the central server; the central server aggregates the gradient to generate global model parameters and distributes them to each terminal. In addition, differential privacy technology is used to add noise to the uploaded gradient to ensure that user data cannot be reversed.

[0027] The present invention will be further described below in conjunction with Examples 1 to 5: Example 1: This example describes in detail the specific methods of data collection, processing and model updating during the construction of the health status dynamic model, and also describes the method of multimodal physiological data preprocessing to make the entire health status dynamic model more complete and accurate.

[0028] Wearable devices and smart environmental sensors play a key role in building a dynamic model of health status. Wearable devices have highly sensitive physiological signal acquisition modules that can continuously and stably collect heart rate, blood pressure, blood oxygen saturation and body movement data. For example, photoelectric sensors are used to measure heart rate and blood oxygen saturation, pressure sensors are used to measure blood pressure, and accelerometers and gyroscopes are used to obtain body movement data. Smart environmental sensors are distributed in various key locations indoors to monitor indoor temperature, humidity, light intensity and air quality index in real time. These sensors transmit the collected data to the data processing terminal through wireless communication technologies (such as Bluetooth, Wi-Fi, etc.).

[0029] The collected multimodal physiological data needs to be preprocessed. Taking heart rate data as an example, artifacts may be introduced due to factors such as movement, affecting the accuracy of the data and subsequent analysis. The heart rate data is processed using wavelet transform technology. Wavelet transform can effectively separate the different frequency components in the signal. By selecting the appropriate wavelet basis and decomposition layer number, motion artifacts can be accurately removed and the true heart rate signal can be retained.

[0030] Blood oxygen saturation data is easily affected by external interference and fluctuates, so Kalman filtering is used for smoothing. Kalman filtering is an algorithm based on linear minimum mean square error estimation. It uses the estimated value of the previous moment and the measured value of the current moment to obtain more accurate and smooth blood oxygen saturation data through continuous iterative updates.

[0031] Body motion data contains rich information, and extracting its frequency domain features helps to better understand the user's motion status. Through Fourier transform and other methods, the energy spectrum density and main frequency components of body motion data are calculated. The energy spectrum density reflects the energy distribution of body motion data at different frequencies, and the main frequency component represents the most important frequency component in the body motion data. These frequency domain features can provide an important basis for subsequent health status analysis.

[0032] After preprocessing, the multimodal physiological data needs to be aligned by timestamp to generate a standardized feature matrix. This ensures that different types of data are consistent in the time dimension, which facilitates subsequent feature extraction and model training.

[0033] When using the Transformer network to extract features from multimodal physiological data, the multi-head attention mechanism in the Transformer network can focus on information at different locations at the same time, thereby better capturing the temporal dependencies in the data. By stacking multiple layers of Transformer layers, more advanced and abstract health status features can be gradually extracted to generate a time-dependent health status vector.

[0034] When the health status vector is spatially fused with the environmental perception data, the environmental perception data is first normalized to make it have the same scale as the health status vector. Then, the two are fused together by splicing and other methods and input into the multi-layer perceptron. The multi-layer perceptron can perform in-depth mining of the fused data through nonlinear transformation of multiple hidden layers to generate a comprehensive health status score.

[0035] The dynamic model of health status is updated through a sliding window mechanism, and the window length is dynamically adjusted according to the user's behavior pattern. For example, for users with relatively fixed daily activity patterns, a longer window length can be set to make full use of historical data for analysis; while for users with large changes in activity patterns, the window length can be appropriately shortened to promptly reflect changes in their health status. By monitoring the user's activity frequency, work and rest time and other behavioral data, combined with machine learning algorithms, the length of the sliding window can be automatically adjusted to ensure that the model can reflect the user's health status in real time and accurately.

[0036] Example 2: When building a health prediction model based on the temporal attention mechanism, the input layer is responsible for receiving historical physiological time series data and real-time environmental perception data. Historical physiological time series data is divided into multiple segments according to the time step. The selection of the time step needs to comprehensively consider the sampling frequency of the data and the actual application requirements. For example, for physiological data with a higher sampling frequency (such as heart rate data, which is sampled multiple times per second), a shorter time step (such as 1 second) can be selected; for data with a lower sampling frequency (such as blood pressure data, which may be measured every few minutes), a longer time step (such as 5 minutes) is selected. This can reduce the amount of calculation while ensuring data integrity.

[0037] The time series embedding layer extracts local features of each time segment through a one-dimensional convolutional network. The convolution kernel in the one-dimensional convolutional network slides on the time dimension and performs convolution operations on each time segment, which can effectively extract local time features. For example, the convolution kernel can capture the changing trend of heart rate in a short period of time, the periodic characteristics of body movement data, etc. At the same time, the time position coding is generated, which can provide the model with the location information of the data in the time series, helping the model to better understand the sequence and time dependency of the data.

[0038] The multi-head attention layer assigns attention weights to local features and temporal position encodings. The multi-head attention mechanism calculates attention weights in parallel through multiple attention heads, and each attention head focuses on a different feature subspace, so that long-term and short-term dependencies can be captured more comprehensively. For example, some attention heads may pay more attention to recent changes in physiological data to capture sudden health risks, while some attention heads focus on long-term trends to predict the development of chronic diseases. By optimizing the attention weights, the model can be more focused on information that has an important impact on health risk prediction.

[0039] The output layer maps the attention weight distribution results through the fully connected network and outputs the probability distribution of the health risk level. The fully connected network can perform nonlinear transformations on the input features and map them to the probability space of the health risk level. During the training process, the weights of the fully connected network are adjusted to make the output of the model as close as possible to the actual health risk level.

[0040] In the process of training the health prediction model, data collection is the key first step. Collect physiological data from wearable devices of at least 1,000 users over a period of years and annotate them with health event labels. These health event labels include abnormal physiological events (such as abnormal heart rate, high or low blood pressure, etc.) and potential health threat events (such as health problems that may be caused by long-term lack of sleep, etc.). Through a large amount of historical data and accurate label annotation, sufficient information is provided for model training.

[0041] Since the number of samples of different health events may be unbalanced, the number of minority health event samples may be small, which will affect the training effect of the model. Time series interpolation and adversarial generative networks are used to expand minority health event samples. Time series interpolation increases the number of samples by inserting new data points between existing samples; adversarial generative networks generate samples similar to real data through adversarial training of generators and discriminators, further expanding minority samples. This can improve the model's ability to identify minority health events and improve overall prediction accuracy.

[0042] The model validation uses five-fold cross validation to evaluate the prediction accuracy. The collected data is divided into five parts, four of which are selected as training sets and one as test sets each time. Through five different divisions and training, the prediction accuracy of the model on different test sets is calculated, and the average value is taken as the final evaluation result. At the same time, the SHAP value is introduced to explain the decision basis of the model. The SHAP value can quantify the influence of each feature on the model output, help users understand why the model makes such a prediction, and improve the interpretability of the model.

[0043] Embodiment 3: When processing environmental perception data, for the air quality index, a segmented quantization method is used to generate air quality level labels. For example, the air quality index is divided into different levels such as excellent (0-50), good (51-100), light pollution (101-150), moderate pollution (151-200), heavy pollution (201-300) and severe pollution (greater than 300), and each level corresponds to a label. In this way, the continuous air quality index can be converted into discrete levels that are easy to understand and process, which is convenient for subsequent analysis and application.

[0044] The day and night patterns of light intensity are identified, and the light adaptability coefficient is generated in combination with the user's work and rest data. The light intensity data is obtained through the light sensor, and the day and night pattern is judged according to the preset light intensity threshold (such as light intensity greater than 500lux during the day and less than 50lux at night). At the same time, the user's work and rest data (such as waking time, sleeping time, etc.) is collected to analyze the user's adaptation in different lighting environments. For example, if the user is used to getting up at 7 o'clock in the morning, and the light intensity is weak at this time, it means that the user has a certain adaptability to low light intensity, and the light adaptability coefficient is increased accordingly; conversely, if the user feels uncomfortable in a strong light environment, the light adaptability coefficient is reduced. The light adaptability coefficient can reflect the user's adaptability to the current light environment and provide a reference for adjusting the health care plan.

[0045] The environmental comfort index is generated by fusing temperature, humidity and air quality level labels through fuzzy logic algorithm. Fuzzy logic algorithm can process information with fuzziness and uncertainty, which is more in line with human subjective feelings about environmental comfort. First, the temperature, humidity and air quality level labels are fuzzified and converted into fuzzy sets. For example, the temperature is divided into fuzzy sets such as cold, cool, comfortable, and hot, and the humidity is divided into fuzzy sets such as dry, moderate, and humid. Then, according to the preset fuzzy rules, these fuzzy sets are inferred and synthesized to generate the environmental comfort index. Fuzzy rules can be set based on a large amount of experimental data and expert experience. For example, when the temperature is comfortable, the humidity is moderate, and the air quality is excellent, the environmental comfort index is high; when the temperature is too high or too low, the humidity is abnormal, and the air quality is poor, the environmental comfort index is low.

[0046] The environmental comfort index is used to adjust the exercise intensity recommendations in the health care plan parameters. When the environmental comfort index is high, it means that the environment is suitable for exercise and the exercise intensity can be appropriately increased; when the environmental comfort index is low, such as in an environment with high temperature, high humidity or severe air pollution, the exercise intensity should be reduced to avoid adverse effects on the body.

[0047] When generating dynamic health intervention strategies, intervention priorities are divided based on health risk levels. For high-priority risks with high probability of physiological abnormalities or high potential health threat indexes, immediate warnings and emergency suggestions are triggered. For example, when the user's heart rate is suddenly too high or blood pressure is out of the normal range, the system immediately issues an alarm and pushes suggestions to the user to seek emergency medical treatment or take emergency measures (such as rest, taking emergency medication, etc.).

[0048] For low-priority risks, a reinforcement learning algorithm is used to generate a long-term health care plan. The reinforcement learning algorithm uses user behavior feedback as a reward signal to optimize the long-term effect of the intervention strategy. For example, when a user follows the diet plan and exercise intensity recommendations given by the system for a period of time, his physical condition improves and his health risk level decreases, the system will give a positive reward; conversely, if the user does not comply with the health care plan, the health risk level does not decrease or even increases, a negative reward will be given. Through continuous rewards and punishments, the reinforcement learning algorithm can gradually find the optimal health care plan and improve the user's health level.

[0049] The adjustment formula for the health care plan parameters is: in, is the intervention intensity at time t, For real-time health risk level, is the environmental comfort index, is an indicator of user behavior compliance. is the weight coefficient. Adjust according to actual conditions. For example, if more attention is paid to the impact of real-time health risks on intervention intensity, the If you want environmental factors to have a greater impact on the intensity of intervention, increase By adjusting these weight coefficients, the adjustment of health care plan parameters can be made more in line with the actual needs of users.

[0050] Example 4: When using the federated learning framework to perform distributed training on user data, each terminal device (such as the user's wearable device, home smart health monitoring device, etc.) undertakes the important task of training the health prediction model sub-model locally. The terminal device stores the user's original data locally, avoiding the privacy leakage risk that may be caused by uploading the data to the cloud.

[0051] The terminal device uses the user data stored locally to train the sub-model. During the training process, only the model gradient is uploaded to the central server instead of uploading the original data directly. The model gradient contains the parameter update direction and amplitude information of the model during the training process. By uploading the gradient, the central server can obtain the training status of the model on each terminal device without involving the user's original data. For example, in one training iteration, the terminal device calculates the gradient of the model parameters based on the local data, packages and encrypts these gradients, and uploads them to the central server.

[0052] The central server is responsible for aggregating the gradients uploaded by each terminal device. After receiving the gradients from multiple terminal devices, the central server uses a suitable aggregation algorithm (such as the average aggregation algorithm) to fuse these gradients and generate global model parameters. The average aggregation algorithm adds the gradients of each terminal device and calculates the average value as the update direction of the global model parameters. In this way, the central server can integrate the training results of each terminal device to obtain a more representative global model.

[0053] After generating the global model parameters, the central server distributes them to each terminal. After receiving the global model parameters, the terminal device applies them to the local sub-model and updates the sub-model parameters. In this way, the sub-model of each terminal device can continuously absorb the advantages of the global model and improve the performance of the model.

[0054] In order to further ensure that user data cannot be reversed, differential privacy technology is used to add noise to the uploaded gradient. Differential privacy technology adds random noise to the gradient, making it difficult for attackers to infer the user's original data from the uploaded gradient. The amount of noise added needs to be adjusted according to the intensity requirements of privacy protection. The greater the added noise, the better the privacy protection effect, but it may have a certain impact on the training accuracy of the model. Therefore, it is necessary to find a balance between privacy protection and model accuracy. For example, through experiments and theoretical analysis, the appropriate amount of noise added can be determined to minimize the impact on the model training accuracy while ensuring the privacy and security of user data.

[0055] Example 5: This example is used to illustrate how to collect and process user behavior feedback, and how to use this feedback to optimize dynamic health intervention strategies to improve the quality and effectiveness of health care services.

[0056] When collecting user behavior feedback, natural language processing is used to parse the user's text evaluation of health care suggestions. After using health care services for a period of time, users may express their own views and opinions on health care suggestions (such as diet plans, exercise intensity suggestions, etc.). Users can submit text evaluations through the comment function of mobile applications, voice input, etc. Natural language processing technology first pre-processes these texts, including removing stop words, word segmentation, and other operations. Then, the sentiment analysis algorithm is used to determine the sentiment tendency of the text, whether it is positive, negative, or neutral. For example, if a user comments that "this diet plan is very suitable for me, and I feel much better recently", it can be judged as a positive evaluation through sentiment analysis; if the user says "the exercise intensity is too high, I can't stick to it at all", it is judged as a negative evaluation. By analyzing a large number of text evaluations, we can understand the user's satisfaction and acceptance of health care suggestions.

[0057] At the same time, the user's emotional state is captured through the camera and microphone to generate an emotional polarity score. The camera can capture the user's facial expressions, such as smiling, frowning, etc., and the microphone can collect the user's voice intonation information. Using computer vision and speech recognition technology, the user's facial expressions and voice intonation are analyzed to determine the user's emotional state, such as happiness, anger, frustration, etc. The emotional state is converted into an emotional polarity score. For example, happiness corresponds to a higher positive score, anger corresponds to a higher negative score, and neutral emotions correspond to a score close to zero. The emotional polarity score can more intuitively reflect the user's emotional changes and provide more comprehensive information for understanding the user's feelings about health care services.

[0058] The text evaluation is combined with the sentiment polarity score to generate a behavior compliance index. The behavior compliance index takes into account the user's evaluation of the health care recommendations and emotional state, and can more accurately measure whether the user is willing to comply with the health care plan. For example, if the user's text evaluation is positive and the sentiment polarity score is positive, it means that the user is relatively satisfied with the health care recommendations and has a high behavior compliance; conversely, if the text evaluation is negative and the sentiment polarity score is negative, the behavior compliance is low.

[0059] Behavioral compliance indicators are used to optimize dynamic health intervention strategies. When the behavioral compliance indicator is low, it means that there may be problems with the current health care plan and it needs to be adjusted. For example, if it is found that many users have low behavioral compliance with exercise intensity recommendations, the exercise intensity may need to be reduced; if users have a poor evaluation of the diet plan, you can consider adjusting the content and combination of the diet plan. By continuously optimizing dynamic health intervention strategies based on behavioral compliance indicators, users' acceptance and satisfaction with health care services can be improved, thereby better achieving health care goals.

[0060] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0061] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A health care data processing method based on intelligent perception, characterized in that: include: Constructing a dynamic health status model, including collecting multimodal physiological data and environmental perception data, and constructing a dynamically updated health status representation model by fusing the multimodal physiological data and environmental perception data; the multimodal physiological data includes heart rate, blood pressure, blood oxygen saturation and body movement data; The environmental sensing data includes indoor temperature, humidity, light intensity and air quality index; Performing health risk prediction, including constructing a health prediction model based on a temporal attention mechanism, inputting historical temporal data of the multimodal physiological data and real-time environmental perception data into the health prediction model, and outputting a health risk level in a future time window; The health risk level includes the probability of physiological abnormality and the potential health threat index; Designing dynamic health intervention strategies, including generating personalized health care recommendations based on the health risk levels output by the health prediction model; The dynamic health intervention strategy includes adjusting health care plan parameters, which include exercise intensity recommendations, diet plans, and rest cycles; Perform privacy protection and model optimization, including using a federated learning framework to perform distributed training on user data, updating parameters of the health prediction model, and ensuring local storage and encrypted transmission of user data.

2. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The construction of the health status dynamic model includes: Collect the multimodal physiological data and environmental perception data in real time through wearable devices and intelligent environmental sensors; Using a Transformer network to extract features from the multimodal physiological data to generate a time-dependent health state vector; spatially fusing the health status vector with environmental perception data to generate a comprehensive health status score through a multi-layer perceptron; The health status dynamic model is updated through a sliding window mechanism, and the window length is dynamically adjusted according to the user behavior pattern.

3. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The health prediction model of the temporal attention mechanism includes: Input layer: receiving historical physiological time series data and real-time environmental perception data, wherein the historical physiological time series data is divided into multiple segments according to time steps; Temporal embedding layer: extracts local features of each time segment through a one-dimensional convolutional network and generates temporal position encoding; Multi-head attention layer: assigns attention weights to the local features and temporal position encodings to capture long-term and short-term dependencies; Output layer: Map the attention weight allocation results through a fully connected network and output the probability distribution of health risk levels.

4. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The processing of the environmental perception data includes: Quantify the air quality index in sections and generate air quality grade labels; Identify the day and night patterns of light intensity and generate a light adaptability coefficient based on the user's daily routine data; The environmental comfort index is generated by fusing temperature, humidity and air quality level labels through fuzzy logic algorithm; The environmental comfort index is used to adjust the exercise intensity recommendation in the health care plan parameters.

5. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The generation of the dynamic health intervention strategy includes: Prioritize interventions based on health risk levels, with high-priority risks triggering immediate alerts and emergency recommendations; For low-priority risks, a reinforcement learning algorithm is used to generate a long-term health care plan. The reinforcement learning algorithm uses user behavior feedback as a reward signal to optimize the long-term effect of the intervention strategy; The adjustment formula of the health care plan parameters is: in, is the intervention intensity at time t, For real-time health risk level, is the environmental comfort index, is an indicator of user behavior compliance. is the weight coefficient.

6. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The implementation of the federated learning framework includes: Each terminal device locally trains a sub-model of the health prediction model and only uploads the model gradient to the central server; The central server aggregates the gradients to generate global model parameters and distributes them to each terminal; Differential privacy technology is used to add noise to uploaded gradients to ensure that user data cannot be reversed.

7. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The preprocessing of the multimodal physiological data includes: Wavelet transform was used to remove motion artifacts from the heart rate data; Perform Kalman filter smoothing on the blood oxygen saturation data; Extract frequency domain features from body motion data, including energy spectrum density and main frequency components; The processed multimodal data are aligned by timestamp to generate a standardized feature matrix.

8. The health care data processing method based on intelligent perception according to claim 3 is characterized in that: The training process of the health prediction model includes: Data collection: Collect physiological data from wearable devices of at least 1,000 users over a period of years and label health events; Feature enhancement: Expand minority health event samples through time series interpolation and adversarial generative networks; Model validation: Five-fold cross validation was used to evaluate the prediction accuracy, and the SHAP value was introduced to explain the basis for model decision making.

9. The health care data processing method based on intelligent perception according to claim 1 is characterized in that: The collection of user behavior feedback includes: Analyze users’ textual comments on health care suggestions through natural language processing; Capture the user's emotional state through the camera and microphone to generate an emotional polarity score; The text evaluation is integrated with the sentiment polarity score to generate a behavioral compliance index for optimizing the dynamic health intervention strategy.

10. A health care data processing system based on intelligent perception, characterized in that: include: Health status dynamic modeling module: used to collect multimodal physiological data and environmental perception data, and build a dynamically updated health status representation model by fusing the multimodal physiological data and environmental perception data. The multimodal physiological data includes heart rate, blood pressure, blood oxygen saturation and body movement data, and the environmental perception data includes indoor temperature, humidity, light intensity and air quality index; Health risk prediction module: used to build a health prediction model based on the temporal attention mechanism, input the historical temporal data of the multimodal physiological data and the real-time environmental perception data into the health prediction model, and output the health risk level in the future time window, wherein the health risk level includes the probability of physiological abnormality and the potential health threat index; Dynamic health intervention strategy module: used to generate personalized health care suggestions according to the health risk level output by the health prediction model, and the dynamic health intervention strategy includes adjusting health care plan parameters, and the health care plan parameters include exercise intensity suggestions, diet plans and rest cycles; Privacy protection and model optimization module: Use the federated learning framework to perform distributed training on user data, update the parameters of the health prediction model, and ensure the local storage and encrypted transmission of user data.

Images