Model training method, system and medium for prediction of acute respiratory infection

Abstract

The present invention relates to the field of acute respiratory infection prediction technology, and in particular to a model training method, system and medium for acute respiratory infection prediction. The method comprises the following steps: collecting raw data of target patients and dividing time windows to obtain time window data; extracting biorhythm features based on the time window data to obtain biorhythm feature data; performing window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data; arranging the important biorhythm time window feature data in time sequence to obtain multi-scale time series feature data of patients; constructing a VAE model based on the multi-scale time series feature data of patients to obtain a VAE model. The present invention achieves a more accurate and personalized prediction of the risk of acute respiratory infection by solving the problem that the existing models do not take into account individual differences and dynamic changes.

Description

Technical Field

[0001] The present invention relates to the technical field of acute respiratory infection prediction, and in particular to a model training method, system and medium for acute respiratory infection prediction. Background Art

[0002] Traditional prediction models for acute respiratory infections are mainly based on statistical learning methods, such as logistic regression, support vector machines, decision trees, etc. Such methods usually use static data such as demographic characteristics, medical history, and lifestyle as input to train a fixed classification or regression model for predicting infection risk. However, this static modeling paradigm cannot fully consider the dynamic evolution of the patient's health status. The occurrence of acute respiratory infection is a complex dynamic process. The patient's physiological indicators, immune function, etc. are constantly changing. It is difficult to fully reflect the progression of the disease by relying solely on static features. Therefore, existing methods have limitations in modeling the dynamic changes of patients.

[0003] In addition, most previous prediction models adopted a group modeling strategy, that is, using all patient data to train a unified model to predict the infection risk of different individuals. This approach ignores the significant individual differences between patients, such as differences in genetic background, previous diseases, drug responsiveness, etc. For complex diseases such as acute respiratory infections where multiple factors interact, group modeling is difficult to take into account the particularities of different patients, and is prone to biased results, affecting the accuracy and practicality of the prediction. Existing methods have limitations in modeling patient personalization. In general, although traditional acute respiratory infection prediction models have made some progress, they are not deep enough in depicting the dynamic changes and individual differences of patients, and it is difficult to fully meet the needs of accurate prediction and personalized prevention and control. Summary of the invention

[0004] Based on this, it is necessary to provide a model training method, system and medium for acute respiratory infection prediction to solve at least one of the above technical problems.

[0005] To achieve the above object, a model training method for predicting acute respiratory infection comprises the following steps:

[0006] Step S1: collect the original data of the target patient and divide it into time windows to obtain time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data; arrange the important biorhythm time window feature data in time sequence to obtain the patient's multi-scale time series feature data;

[0007] Step S2: construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain clustering data of health trajectories of the patient; generate individual health trajectories of the patient based on the clustering data of health trajectories of the patient and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient;

[0008] Step S3: Acquire the target patient's regional environmental data through the environmental monitoring station to obtain the patient's regional environmental data; perform time dimension alignment according to the patient's regional environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain the patient's environmental sensitivity spectrum;

[0009] Step S4: constructing a dynamic Bayesian network model based on the multi-scale time series feature data of the patient, the individual health trajectory data of the patient, and the environmental sensitivity spectrum of the patient to obtain a dynamic Bayesian network model; generating a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network;

[0010] Step S5: collect real-time data of patients with acute respiratory infection to obtain real-time data of patients; input the real-time data of patients into a personalized dynamic Bayesian network, update the dynamic Bayesian network, and obtain a dynamic Bayesian update network; predict the risk of future acute respiratory infection based on the dynamic Bayesian update network and push personalized acute respiratory infection warning to obtain personalized acute respiratory infection warning data.

[0011] The present invention collects, processes and integrates the electronic medical record data and physiological index data of patients, and extracts biorhythm features to construct multi-scale time series feature data of patients containing rich information, providing a high-quality data basis for subsequent model training and prediction. The VAE model is used to learn and reduce the multi-scale time series feature data of patients, and through trajectory clustering analysis and future trajectory prediction, personalized health trajectory data that can reflect the law of changes in the individual health status of patients is generated, providing more accurate individualized information for subsequent risk prediction. By analyzing the correlation between environmental data and individual health trajectory data of patients, environmental factors that have a significant impact on the health of patients are identified, and a personalized environmental sensitivity spectrum that can reflect the sensitivity of patients to different environmental factors is constructed, providing important environmental information for a comprehensive assessment of the risk of acute respiratory infections in patients. By integrating multi-scale time series feature data of patients, individual health trajectory data and environmental sensitivity spectrum, a personalized dynamic Bayesian network model that can capture individual differences, dynamic changes and environmental influences of patients is constructed using a Bayesian network, providing an effective tool for more accurate prediction of the risk of acute respiratory infections. By collecting patient data in real time and updating the personalized dynamic Bayesian network model, real-time monitoring, dynamic evaluation and personalized early warning of the risk of acute respiratory infection in patients are realized, providing more timely, more accurate and more personalized services for the health management of patients, which helps to reduce the risk of disease. Therefore, the present invention provides a model training method for acute respiratory infection prediction. In order to solve the deficiencies of traditional model training methods for acute respiratory infection prediction in depicting the dynamic changes and individual differences of patients, a series of designs such as time dimension modeling, multi-scale feature fusion, individualized representation learning, knowledge-driven optimization, and dynamic update early warning are used to capture the dynamic changes and individual differences of patients. Comprehensive consideration and modeling are carried out, which makes up for the shortcomings of traditional methods, effectively solves the problem that the existing model does not consider the individual differences and dynamic changes, and realizes more accurate and personalized prediction of acute respiratory infection risks, providing a more effective tool for disease prevention.

[0012] Preferably, the present invention also provides a model training system for acute respiratory infection prediction, which is used to execute the model training method for acute respiratory infection prediction as described above, and the model training system for acute respiratory infection prediction comprises:

[0013] The multi-scale time series construction module is used to collect the original data of the target patient and divide the time window to obtain the time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain the important biorhythm time window feature data; arrange the important biorhythm time window feature data in time order to obtain the multi-scale time series feature data of the patient;

[0014] The individual health trajectory modeling module is used to construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain cluster data of the patient's health trajectory; generate individual health trajectories of the patient based on the cluster data of the patient's health trajectory and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient;

[0015] The environmental sensitivity analysis module is used to obtain the patient area environmental data of the target patient through the environmental detection station to obtain the patient area environmental data; perform time dimension alignment according to the patient area environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain a patient environmental sensitivity spectrum;

[0016] A dynamic Bayesian network construction module is used to construct a dynamic Bayesian network model based on the patient's multi-scale time series feature data, the patient's individual health trajectory data, and the patient's environmental sensitivity spectrum to obtain a dynamic Bayesian network model; and to generate a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network;

[0017] The real-time risk prediction and early warning module is used to collect real-time data of patients with acute respiratory infections to obtain real-time data of patients; input the real-time data of patients into a personalized dynamic Bayesian network, update the dynamic Bayesian network, and obtain a dynamic Bayesian update network; predict future acute respiratory infection risks based on the dynamic Bayesian update network and push personalized acute respiratory infection early warnings to obtain personalized acute respiratory infection early warning data.

[0018] Preferably, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed, implements the model training method for predicting acute respiratory tract infection as described above.

[0019] The present invention collects, processes and integrates the electronic medical record data and physiological index data of patients, and extracts biorhythm features to construct multi-scale time series feature data of patients containing rich information, providing a high-quality data basis for subsequent model training and prediction. The VAE model is used to learn and reduce the multi-scale time series feature data of patients, and through trajectory clustering analysis and future trajectory prediction, personalized health trajectory data that can reflect the law of changes in the individual health status of patients is generated, providing more accurate individualized information for subsequent risk prediction. By analyzing the correlation between environmental data and individual health trajectory data of patients, environmental factors that have a significant impact on the health of patients are identified, and a personalized environmental sensitivity spectrum that can reflect the sensitivity of patients to different environmental factors is constructed, providing important environmental information for a comprehensive assessment of the risk of acute respiratory infections in patients. By integrating multi-scale time series feature data of patients, individual health trajectory data and environmental sensitivity spectrum, a personalized dynamic Bayesian network model that can capture individual differences, dynamic changes and environmental influences of patients is constructed using a Bayesian network, providing an effective tool for more accurate prediction of the risk of acute respiratory infections. By collecting patient data in real time and updating the personalized dynamic Bayesian network model, real-time monitoring, dynamic evaluation and personalized early warning of the risk of acute respiratory infection in patients are realized, providing more timely, more accurate and more personalized services for the health management of patients, which helps to reduce the risk of disease. Therefore, the present invention provides a model training method for acute respiratory infection prediction. In order to solve the deficiencies of traditional model training methods for acute respiratory infection prediction in depicting the dynamic changes and individual differences of patients, a series of designs such as time dimension modeling, multi-scale feature fusion, individualized representation learning, knowledge-driven optimization, and dynamic update early warning are used to capture the dynamic changes and individual differences of patients. Comprehensive consideration and modeling are carried out, which makes up for the shortcomings of traditional methods, effectively solves the problem that the existing model does not consider the individual differences and dynamic changes, and realizes more accurate and personalized prediction of acute respiratory infection risks, providing a more effective tool for disease prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of the steps of a model training method for predicting acute respiratory infection;

[0021] Figure 2 for Figure 1 Detailed implementation steps of step S2 in the flowchart;

[0022] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0023] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0024] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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 technicians in this field without creative work are within the scope of protection of the present invention.

[0025] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0026] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0027] To achieve this, please refer to Figures 1 to 3 , a model training method for acute respiratory infection prediction, comprising the following steps:

[0028] Step S1: collect the original data of the target patient and divide it into time windows to obtain time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data; arrange the important biorhythm time window feature data in time sequence to obtain the patient's multi-scale time series feature data;

[0029] Step S2: construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain clustering data of health trajectories of the patient; generate individual health trajectories of the patient based on the clustering data of health trajectories of the patient and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient;

[0030] Step S3: Acquire the target patient's regional environmental data through the environmental monitoring station to obtain the patient's regional environmental data; perform time dimension alignment according to the patient's regional environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain the patient's environmental sensitivity spectrum;

[0031] Step S4: constructing a dynamic Bayesian network model based on the multi-scale time series feature data of the patient, the individual health trajectory data of the patient, and the environmental sensitivity spectrum of the patient to obtain a dynamic Bayesian network model; generating a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network;

[0032] Step S5: collect real-time data of patients with acute respiratory infection to obtain real-time data of patients; input the real-time data of patients into a personalized dynamic Bayesian network, update the dynamic Bayesian network, and obtain a dynamic Bayesian update network; predict the risk of future acute respiratory infection based on the dynamic Bayesian update network and push personalized acute respiratory infection warning to obtain personalized acute respiratory infection warning data.

[0033] The present invention collects, processes and integrates the electronic medical record data and physiological index data of patients, and extracts biorhythm features to construct multi-scale time series feature data of patients containing rich information, providing a high-quality data basis for subsequent model training and prediction. The VAE model is used to learn and reduce the multi-scale time series feature data of patients, and through trajectory clustering analysis and future trajectory prediction, personalized health trajectory data that can reflect the law of changes in the individual health status of patients is generated, providing more accurate individualized information for subsequent risk prediction. By analyzing the correlation between environmental data and individual health trajectory data of patients, environmental factors that have a significant impact on the health of patients are identified, and a personalized environmental sensitivity spectrum that can reflect the sensitivity of patients to different environmental factors is constructed, providing important environmental information for a comprehensive assessment of the risk of acute respiratory infections in patients. By integrating multi-scale time series feature data of patients, individual health trajectory data and environmental sensitivity spectrum, a personalized dynamic Bayesian network model that can capture individual differences, dynamic changes and environmental influences of patients is constructed using a Bayesian network, providing an effective tool for more accurate prediction of the risk of acute respiratory infections. By collecting patient data in real time and updating the personalized dynamic Bayesian network model, real-time monitoring, dynamic evaluation and personalized early warning of the risk of acute respiratory infection in patients are realized, providing more timely, more accurate and more personalized services for the health management of patients, which helps to reduce the risk of disease. Therefore, the present invention provides a model training method for acute respiratory infection prediction. In order to solve the deficiencies of traditional model training methods for acute respiratory infection prediction in depicting the dynamic changes and individual differences of patients, a series of designs such as time dimension modeling, multi-scale feature fusion, individualized representation learning, knowledge-driven optimization, and dynamic update early warning are used to capture the dynamic changes and individual differences of patients. Comprehensive consideration and modeling are carried out, which makes up for the shortcomings of traditional methods, effectively solves the problem that the existing model does not consider the individual differences and dynamic changes, and realizes more accurate and personalized prediction of acute respiratory infection risks, providing a more effective tool for disease prevention.

[0034] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of the model training method for predicting acute respiratory tract infection of the present invention. In this example, the model training method for predicting acute respiratory tract infection includes the following steps:

[0035] Step S1: collect the original data of the target patient and divide it into time windows to obtain time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data; arrange the important biorhythm time window feature data in time sequence to obtain the patient's multi-scale time series feature data;

[0036] In the embodiment of the present invention, the electronic medical record data of the target patient and the physiological index data collected by the wearable device are collected, and the data is preprocessed and divided according to the time window. The periodicity of the physiological indexes in the time window data is analyzed, the biorhythm features are extracted, and the features related to acute respiratory tract infection are screened, and finally the multi-scale time series feature data of the patient is constructed.

[0037] Step S2: construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain clustering data of health trajectories of the patient; generate individual health trajectories of the patient based on the clustering data of health trajectories of the patient and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient;

[0038] In the embodiment of the present invention, the variational autoencoder (VAE) model is trained using the multi-scale time series feature data of the patient to learn the low-dimensional latent variable representation of the data. The individualized health trajectory data of the patient is generated by smoothing the latent variable sequence, performing trajectory clustering analysis and future trajectory prediction.

[0039] Step S3: Acquire the target patient's regional environmental data through the environmental monitoring station to obtain the patient's regional environmental data; perform time dimension alignment according to the patient's regional environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain the patient's environmental sensitivity spectrum;

[0040] In the embodiment of the present invention, the environmental monitoring data of the patient's residence is obtained and aligned with the patient's individual health trajectory data in the time dimension. The time delay of the impact of environmental factors on health is analyzed, and a time delay-similarity matrix is ​​constructed. Based on this, the environmental factor sensitivity score is calculated, and finally the patient's environmental sensitivity spectrum is constructed.

[0041] Step S4: constructing a dynamic Bayesian network model based on the multi-scale time series feature data of the patient, the individual health trajectory data of the patient, and the environmental sensitivity spectrum of the patient to obtain a dynamic Bayesian network model; generating a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network;

[0042] In the embodiment of the present invention, the multi-scale time series feature data, individual health trajectory data and environmental sensitivity spectrum of the patient are integrated to construct the multi-dimensional data of the patient. The initial network structure is constructed using the Bayesian network structure learning algorithm and modified in combination with the knowledge of medical experts. The parameterized network model is obtained through parameter learning and dynamically expanded according to the individual data of the patient to generate a personalized dynamic Bayesian network.

[0043] Step S5: real-time data collection of patients with acute respiratory tract infection is performed to obtain real-time data of the patients; the real-time data of the patients is input into a personalized dynamic Bayesian network, and the dynamic Bayesian network is updated to obtain a dynamic Bayesian update network; the risk of acute respiratory tract infection in the future is predicted according to the dynamic Bayesian update network, and personalized acute respiratory tract infection warning is pushed to obtain personalized acute respiratory tract infection warning data;

[0044] In the embodiment of the present invention, the patient's physiological indicators, environment, behavior and other data are collected in real time and input into a personalized dynamic Bayesian network for model updating. The updated network is used to predict the patient's risk probability of acute respiratory infection in the future, and personalized acute respiratory infection warning push is performed based on the preset risk level threshold and the patient's real-time status.

[0045] Preferably, step S1 comprises the following steps:

[0046] Step S11: obtaining electronic medical record data of the patient through the medical system; obtaining physiological index data of the patient through the wearable device; performing data preprocessing on the electronic medical record data of the patient and the physiological index data of the patient to obtain preprocessed patient data; dividing the preprocessed patient data into time windows to obtain time window data;

[0047] Step S12: extracting biorhythm features according to the time window data to obtain biorhythm feature data; performing window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data;

[0048] Step S13: performing multi-scale feature fusion on the important biological rhythm time window feature data to obtain multi-scale feature data of the patient;

[0049] Step S14: performing time dimension feature enhancement on the multi-scale feature data of the patient to obtain multi-scale feature enhanced data of the patient;

[0050] Step S15: Arrange the patient's multi-scale feature enhancement data in time sequence to obtain the patient's multi-scale time series feature data.

[0051] The present invention can effectively collect and organize the patient's electronic medical records and physiological index information by integrating patient data from medical systems and wearable devices, and performing preprocessing and time window division, so as to provide a high-quality and structured data basis for subsequent feature extraction and model training. By analyzing the periodic laws of physiological indicators, extracting biorhythm features, and screening features related to acute respiratory infections, it is possible to capture the dynamic changes in the patient's physiological state, identify potential risk factors, and improve the accuracy and sensitivity of model prediction. Multi-scale fusion of important biorhythm time window features with static features in the patient's electronic medical record data can more comprehensively characterize the individual characteristics of the patient, combine the patient's static information with dynamic change trends, and improve the model's ability to express individual differences. The time dimension feature enhancement of the patient's multi-scale feature data, such as using a sliding window method, can capture the dynamic change trend of the patient's health status in the time dimension, better reflect the occurrence and development of acute respiratory infections, and improve the model's learning ability for time series data. The multi-scale feature enhancement data of patients are arranged in chronological order to construct the multi-scale time series feature data of patients, which provides a data basis for the subsequent training and prediction of time series models such as recurrent neural networks, helps the model capture time dependencies and improves the accuracy of predictions.

[0052] In an embodiment of the present invention, first, the electronic medical record data of the target patient is retrieved from the hospital medical system database, including the patient's age, gender, medical history, medication records and other information. At the same time, the patient's heart rate, body temperature, sleep and other physiological index data are collected through the wearable device worn by the patient. The electronic medical record data and the physiological index data are integrated, and preprocessing operations such as data cleaning, missing value processing, and standardization are performed to obtain preprocessed patient data. Finally, according to the preset time window size (for example, 24 hours as a time window), the preprocessed patient data is divided to obtain a series of time window data. For each time window data, the physiological index sequence (such as heart rate, body temperature, etc.) contained therein is analyzed, and the significant periodic components of the physiological index are identified by Fourier transform and other methods, and the main cycle length is determined (for example, the main cycle length of the heart rate is 24 hours). According to the main cycle length, the physiological index sequence is segmented and feature extracted, such as calculating the statistical features such as the average value, maximum value, minimum value, and variance in each cycle to obtain the biological rhythm feature data. Next, feature selection methods such as random forests are used to screen out biorhythm time window features that are highly correlated with the prediction of acute respiratory tract infection, and obtain important biorhythm time window feature data. The important biorhythm time window feature data is integrated with the static features (such as age, gender, past medical history, etc.) in the patient's electronic medical record data. Multi-scale feature fusion can be performed using methods such as feature splicing and feature crossing to obtain feature data containing multi-scale information of patients. In order to enhance the information expression ability in the time dimension, the time dimension feature enhancement of the patient's multi-scale feature data is performed. For example, the sliding window method can be used to integrate the multi-scale feature data over a period of time in the past to construct a new feature vector, thereby capturing the dynamic change trend of the patient in the time dimension. Finally, the patient's multi-scale feature enhancement data is arranged in chronological order to construct the patient's multi-scale time series feature data, so as to prepare data for subsequent model training.

[0053] Preferably, step S12 comprises the following steps:

[0054] Step S121: extracting physiological indicators from the time window data to obtain a physiological indicator sequence;

[0055] Step S122: performing periodic analysis on the physiological index sequence to obtain significant periodic component data; determining the main period length according to the significant periodic component data to obtain the main period data;

[0056] Step S123: extracting biorhythm features from the physiological indicator sequence according to the main cycle data to obtain biorhythm feature data;

[0057] Step S124: constructing biorhythm time window features for the biorhythm feature data to obtain biorhythm time window feature data;

[0058] Step S125: Filter the biological rhythm time window feature data by window feature to obtain important biological rhythm time window feature data.

[0059] The present invention can convert the original and scattered physiological indicator data into the form of time series by extracting the physiological indicator sequence from the time window data, so as to provide a basis for subsequent periodic analysis and feature extraction. By performing periodic analysis on the physiological indicator sequence, identifying the significant periodic components, and determining the main period length, the law of human physiological rhythm, such as circadian rhythm, respiratory rhythm, etc., can be captured, and a more refined time scale can be provided for subsequent feature extraction. The physiological indicator sequence is segmented according to the main period data, and the statistical features of each period are extracted, so that more representative biological rhythm features can be constructed, such as the daily mean value of heart rate, night variability, etc., which can better reflect the health status of the patient. Combining the biological rhythm features with the time window information, constructing the biological rhythm time window features, the physiological rhythm information can be integrated with the time background information, and the health status of the patient can be described more comprehensively. The biological rhythm time window features are screened, and the features highly related to acute respiratory tract infection are retained, so that redundant information can be removed, the efficiency and prediction performance of the model are improved, and the interpretability of the model is also enhanced.

[0060] In an embodiment of the present invention, pre-set physiological indicator data is extracted from each time window data to form time series data corresponding to each indicator. For example, the heart rate value recorded every hour in a certain time window is extracted to form a heart rate time series in the time window. Periodic analysis is performed on each physiological indicator sequence. The time series data can be converted to the frequency domain using methods such as fast Fourier transform (FFT), and statistically significant periodic components are identified, and their corresponding frequency and amplitude information are recorded. According to a preset significance threshold, significant periodic component data is screened out. According to the component with the highest frequency in the significant periodic component data, the main cycle length of the physiological indicator is determined, for example, the main cycle length of the heart rate is usually 24 hours. According to the main cycle length, each physiological indicator sequence is divided into multiple cycle segments. For each cycle segment, its statistical features are extracted to construct biorhythm feature data. The biorhythm feature data is arranged in the order of the time window, and combined with the tag information of the time window (for example, which day of the week the time window is, which time period of the day, etc.), the biorhythm time window feature data is constructed. Using machine learning algorithms and combining the patient's acute respiratory infection occurrence as a label, the biorhythm time window feature data is sorted by feature importance, and features that are highly correlated with the risk of acute respiratory infection are screened out to obtain important biorhythm time window feature data.

[0061] Preferably, step S2 comprises the following steps:

[0062] Step S21: dividing the patient's multi-scale time series feature data into data sets to obtain a training set and a verification set; performing sliding window data augmentation and random noise addition on the training set and the verification set to obtain an enhanced training set and an enhanced verification set;

[0063] Step S22: constructing a VAE model according to the multi-scale time series feature data of the patient to obtain a VAE model framework;

[0064] Step S23: Using the enhanced training set to perform model training on the VAE model framework and using the enhanced validation set to perform model evaluation to obtain a VAE model;

[0065] Step S24: Using the VAE model to extract latent variables from the patient's multi-scale time series feature data to obtain the patient's latent variable sequence;

[0066] Step S25: Performing latent variable space visualization on the patient's latent variable sequence to obtain a patient health status distribution diagram; performing time smoothing processing on the patient's latent variable sequence to obtain a patient's latent variable smoothed sequence;

[0067] Step S26: performing trajectory similarity calculation on the patient's latent variable smoothing sequence according to the patient's health status distribution diagram to obtain the patient's health trajectory similarity matrix; performing trajectory clustering analysis on the patient's health trajectory similarity matrix to obtain the patient's health trajectory clustering data;

[0068] Step S27: predicting the future trajectory according to the patient's health trajectory clustering data and the patient's latent variable smoothing sequence to obtain the patient's future health trajectory data;

[0069] Step S28: Generate the individual health trajectory of the patient according to the smoothed sequence of the patient's latent variables and the patient's future health trajectory data to obtain the individual health trajectory data of the patient.

[0070] The present invention can effectively expand the amount of training data, improve the generalization ability of the model, enhance the robustness of the model to noise, prevent overfitting, and thus improve the prediction performance of the model by dividing the data set into a training set and a validation set, and performing sliding window data enhancement and random noise addition. The VAE model framework is constructed based on the multi-scale time series feature data of the patient, and the complex nonlinear relationships in the data can be learned using deep learning methods, and the high-dimensional time series data can be mapped to a low-dimensional latent variable space to achieve data dimension reduction and feature extraction. The VAE model framework is trained using an enhanced training set, and the model is evaluated using an enhanced validation set, which can optimize the model parameters, improve the model's fitting ability and generalization ability, select the model with the best performance, and ensure the prediction effect of the model. Using the trained VAE model to extract latent variables from the multi-scale time series feature data of the patient can convert the original data into a more representative and informative low-dimensional vector, retain the key information in the data, and facilitate subsequent analysis and modeling. Visualizing and time-smoothing the patient's latent variable sequence can intuitively observe the distribution of the patient's health status and eliminate the influence of high-frequency noise, making the latent variable sequence smoother and easier to perform subsequent trajectory analysis. Trajectory similarity calculation and cluster analysis of the patient's latent variable smooth sequence based on the patient's health status distribution diagram can classify patients with similar health trajectories into the same category, identify different health patterns, and provide a basis for personalized prediction. Future trajectory prediction based on the patient's health trajectory clustering data and the patient's latent variable smooth sequence can predict the health trend for a period of time in the future based on the patient's historical health status, providing a reference for early warning of the disease. Combining the patient's current latent variable smooth sequence and the predicted future health trajectory data to generate complete patient individual health trajectory data can more comprehensively describe the patient's health status and provide a personalized data basis for subsequent acute respiratory infection risk prediction.

[0071] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0072] Step S21: dividing the patient's multi-scale time series feature data into data sets to obtain a training set and a verification set; performing sliding window data augmentation and random noise addition on the training set and the verification set to obtain an enhanced training set and an enhanced verification set;

[0073] In the embodiment of the present invention, the multi-scale time series feature data of the patient obtained in step S15 is divided into a training set and a validation set according to a certain ratio (e.g., 8:2). In order to expand the training data and improve the generalization ability of the model, the training set and the validation set are respectively enhanced by sliding window data, that is, the time series is intercepted by a sliding window to generate multiple new time series samples. At the same time, a certain proportion of random noise is added to the enhanced data to simulate the noise interference in the real data, so as to obtain an enhanced training set and an enhanced validation set.

[0074] Step S22: constructing a VAE model according to the multi-scale time series feature data of the patient to obtain a VAE model framework;

[0075] In the embodiment of the present invention, according to the dimension and data distribution characteristics of the multi-scale time series feature data of the patient, a variational autoencoder (VAE) model framework is constructed, including an encoder, a decoder, a loss function, etc. The encoder is used to map the high-dimensional time series data to a low-dimensional latent variable space, and the decoder restores the vector in the latent variable space to time series data.

[0076] Step S23: Using the enhanced training set to perform model training on the VAE model framework and using the enhanced validation set to perform model evaluation to obtain a VAE model;

[0077] In an embodiment of the present invention, an enhanced training set is used to train the VAE model framework, and the model parameters are learned by minimizing the reconstruction loss and the KL divergence loss. The performance of the model is evaluated using an enhanced validation set, such as calculating indicators such as the reconstruction error and the latent variable distribution, and the model with the best performance is selected as the final VAE model.

[0078] Step S24: Using the VAE model to extract latent variables from the patient's multi-scale time series feature data to obtain the patient's latent variable sequence;

[0079] In the embodiment of the present invention, the trained VAE model is used to process the multi-scale time series feature data of the patient, and the time series data is input into the encoder to obtain the corresponding low-dimensional latent variable representation to form the patient latent variable sequence.

[0080] Step S25: Performing latent variable space visualization on the patient's latent variable sequence to obtain a patient health status distribution diagram; performing time smoothing processing on the patient's latent variable sequence to obtain a patient's latent variable smoothed sequence;

[0081] In the embodiment of the present invention, the patient latent variable sequence is mapped to a two-dimensional plane for visualization to observe the distribution of the patient's health status. For example, it can be observed whether the patient's health status is clustered together, or whether there are obvious abnormal points, etc. At the same time, a time smoothing method such as moving average is applied to the patient latent variable sequence to eliminate the influence of high-frequency noise and obtain a smoother patient latent variable smoothed sequence.

[0082] Step S26: performing trajectory similarity calculation on the patient's latent variable smoothing sequence according to the patient's health status distribution diagram to obtain the patient's health trajectory similarity matrix; performing trajectory clustering analysis on the patient's health trajectory similarity matrix to obtain the patient's health trajectory clustering data;

[0083] In an embodiment of the present invention, the similarity between the smoothed sequences of latent variables of different patients is calculated based on the patient health status distribution map. For example, the dynamic time warping (DTW) algorithm can be used to calculate the distance between time series and construct a patient health trajectory similarity matrix. Then, the similarity matrix is ​​clustered using methods such as hierarchical clustering to classify patients with similar health trajectories into the same category, thereby obtaining patient health trajectory clustering data.

[0084] Step S27: predicting the future trajectory according to the patient's health trajectory clustering data and the patient's latent variable smoothing sequence to obtain the patient's future health trajectory data;

[0085] In an embodiment of the present invention, the patient's health trajectory for a period of time in the future is predicted based on the patient's health trajectory clustering data and the patient's current hidden variable smooth sequence. For example, a model such as a recurrent neural network (RNN) can be used to take the patient's historical hidden variable smooth sequence as input to predict the hidden variables at future moments, thereby obtaining the patient's future health trajectory data.

[0086] Step S28: generating the individual health trajectory of the patient according to the smoothed sequence of the patient's latent variables and the patient's future health trajectory data to obtain the individual health trajectory data of the patient;

[0087] In the embodiment of the present invention, the patient's current hidden variable smoothed sequence and the patient's future health trajectory data are combined to construct complete patient individual health trajectory data, providing a personalized data basis for subsequent acute respiratory infection risk prediction.

[0088] Preferably, step S3 comprises the following steps:

[0089] Step S31: acquiring the patient area environmental data of the target patient through the environmental detection station to obtain the patient area environmental data; performing time alignment processing on the patient area environmental data according to the patient's individual health trajectory data to obtain time-aligned environmental data;

[0090] Step S32: extracting infection risk-related environmental features from the time-aligned environmental data to obtain an environmental feature time series; performing multi-scale decomposition on the environmental feature time series to obtain a multi-scale environmental feature time series;

[0091] Step S33: aligning the multi-scale environmental feature time series and the patient's individual health trajectory data in time dimension to obtain health and environment alignment data;

[0092] Step S34: constructing a time delay-similarity matrix according to the time-aligned environment data and the individual health trajectory data of the patient to obtain a time delay-similarity matrix; extracting a time delay parameter according to the time delay-similarity matrix to obtain a time delay parameter;

[0093] Step S35: constructing a health-environment similarity matrix according to the time delay parameter and the time delay-similarity matrix to obtain a health-environment similarity matrix;

[0094] Step S36: Calculate the sensitivity score of the health and environment alignment data according to the health-environment similarity matrix to obtain environmental factor sensitivity score data; perform sensitivity feature screening according to the environmental factor sensitivity score data to obtain key environmental factor data;

[0095] Step S37: construct a patient environmental sensitivity spectrum based on the key environmental factor data and the environmental factor sensitivity score data to obtain the patient environmental sensitivity spectrum.

[0096] The present invention can associate environmental factors with the patient's health status in the time dimension by acquiring environmental data in the patient's area and performing time alignment according to the patient's individual health trajectory data, thereby providing a basis for subsequent analysis of the impact of environmental factors on health. Extracting environmental features related to the risk of respiratory infection and performing multi-scale decomposition can capture the changing laws of environmental factors at different time scales, such as short-term fluctuations and long-term trends, and provide more detailed information for subsequent analysis. Aligning the multi-scale environmental feature time series and the patient's individual health trajectory data in the time dimension can establish a time correspondence between environmental factors and health status, and provide a more accurate time window for subsequent analysis of the impact of environmental factors on health. By constructing a time delay-similarity matrix and extracting time delay parameters, the lag effect of environmental factors on health can be identified. For example, after the PM2.5 concentration increases, the patient will not experience respiratory symptoms until a period of time later. Adjusting the time delay-similarity matrix according to the time delay parameter and constructing a health-environment similarity matrix can more accurately reflect the degree of association between environmental factors and health status after considering the time delay. By calculating the environmental factor sensitivity score based on the health-environment similarity matrix and screening key environmental factors, we can identify environmental factors that have a greater impact on patients' health. For example, some patients are more sensitive to changes in PM2.5 concentration, while others are more sensitive to changes in temperature. Constructing a patient environmental sensitivity spectrum can intuitively show the patient's sensitivity to different environmental factors and provide a reference for personalized health management. For example, for patients who are sensitive to PM2.5 concentration, they can be reminded to take protective measures when air pollution is severe.

[0097] As an example of the present invention, refer to Figure 3 As shown, in this example, step S3 includes:

[0098] Step S31: acquiring the patient area environmental data of the target patient through the environmental detection station to obtain the patient area environmental data; performing time alignment processing on the patient area environmental data according to the patient's individual health trajectory data to obtain time-aligned environmental data;

[0099] In the embodiment of the present invention, according to the residence information of the target patient, the environmental data of the area is obtained from the environmental monitoring station, such as PM2.5 concentration, temperature, humidity, air pressure, wind speed, etc. According to the timestamp information of the patient's individual health trajectory data, the obtained environmental data is time-aligned to ensure the consistency of the environmental data and health data in the time dimension, and obtain time-aligned environmental data.

[0100] Step S32: extracting infection risk-related environmental features from the time-aligned environmental data to obtain an environmental feature time series; performing multi-scale decomposition on the environmental feature time series to obtain a multi-scale environmental feature time series;

[0101] In the embodiment of the present invention, environmental characteristics related to the risk of respiratory infection, such as PM2.5 concentration, temperature, humidity, etc., are extracted from the time-aligned environmental data to construct an environmental characteristic time series. Then, the environmental characteristic time series is decomposed using a multi-scale decomposition method such as wavelet transform to obtain environmental characteristic time series at different time scales, so as to analyze the impact of environmental factors at different time scales on patient health.

[0102] Step S33: aligning the multi-scale environmental feature time series and the patient's individual health trajectory data in time dimension to obtain health and environment alignment data;

[0103] In an embodiment of the present invention, the multi-scale environmental feature time series and the patient's individual health trajectory data are aligned in the time dimension according to the timestamp information to ensure the one-to-one correspondence between the environmental data and the health data at the time point, thereby obtaining health and environment aligned data.

[0104] Step S34: constructing a time delay-similarity matrix according to the time-aligned environment data and the individual health trajectory data of the patient to obtain a time delay-similarity matrix; extracting a time delay parameter according to the time delay-similarity matrix to obtain a time delay parameter;

[0105] In an embodiment of the present invention, considering that there is a time delay in the impact of environmental factors on human health, a time delay-similarity matrix is ​​constructed. The rows of the matrix represent different time delays, the columns represent different environmental features, and the matrix elements represent the similarity between the environmental feature time series and the patient's individual health trajectory data under the time delay. The dynamic time warping (DTW) algorithm can be used to calculate the similarity between time series. According to the time delay-similarity matrix, the maximum similarity corresponding to each environmental feature and its corresponding time delay are found as the time delay parameter of the environmental feature.

[0106] Step S35: constructing a health-environment similarity matrix according to the time delay parameter and the time delay-similarity matrix to obtain a health-environment similarity matrix;

[0107] In the embodiment of the present invention, the time delay-similarity matrix is ​​adjusted to a health-environment similarity matrix according to the time delay parameter. The rows of the matrix represent different environmental features, the columns represent the time points of the individual health trajectory data of the patient, and the matrix elements represent the similarity between the environmental feature and the patient's health status at that time point after considering the time delay.

[0108] Step S36: Calculate the sensitivity score of the health and environment alignment data according to the health-environment similarity matrix to obtain environmental factor sensitivity score data; perform sensitivity feature screening according to the environmental factor sensitivity score data to obtain key environmental factor data;

[0109] In the embodiment of the present invention, the degree of influence of each environmental feature on the individual health status of the patient is calculated according to the health-environment similarity matrix, that is, the environmental factor sensitivity score. For example, the average value of the corresponding row of each environmental feature in the health-environment similarity matrix can be calculated as the sensitivity score of the environmental feature. Then, according to the preset threshold or importance ranking, the key environmental factors that have a greater impact on the patient's health status are screened out.

[0110] Step S37: constructing a patient environmental sensitivity spectrum based on the key environmental factor data and the environmental factor sensitivity score data to obtain the patient environmental sensitivity spectrum;

[0111] In an embodiment of the present invention, a patient environmental sensitivity spectrum is constructed based on key environmental factor data and environmental factor sensitivity score data. The sensitivity spectrum can intuitively show the patient's sensitivity to different environmental factors, for example, it can be visualized in the form of a bar graph, a heat map, etc., to provide personalized environmental factor analysis results for subsequent acute respiratory infection risk prediction.

[0112] Preferably, step S34 includes the following steps:

[0113] Step S341: defining a time delay range according to the time alignment environment data to obtain time delay range data; generating a time delay sequence according to the time delay range data to obtain a time delay sequence;

[0114] Step S342: constructing a delayed environment feature sequence set for the environment feature time series according to the time delay sequence to obtain a delayed environment feature sequence set;

[0115] Step S343: Calculate the time delay similarity of the delay environment feature sequence set and the patient's individual health trajectory data to obtain a time delay similarity value; construct a time delay-similarity matrix according to the time delay similarity value and the delay environment feature sequence set to obtain a time delay-similarity matrix;

[0116] Step S344: extracting time delay parameters from the time delay-similarity matrix according to the delay environment feature sequence set to obtain time delay parameters.

[0117] By defining the time delay range and generating a time delay sequence, the present invention can systematically consider the lag effect of environmental factors on health, covering from immediate impact to potential impact within a longer time span, and provide more comprehensive time dimension information for subsequent analysis. According to the time delay sequence, a delayed environmental feature sequence set is constructed, and the environmental feature time series can be copied and translated according to different time delays to generate a series of alternative environmental data that take into account different time delays, providing a data basis for the subsequent calculation of the time delay-similarity matrix. The similarity between the delayed environmental feature sequence set and the individual health trajectory data of the patient is calculated, and a time delay-similarity matrix is ​​constructed, which can intuitively show the degree of correlation between environmental factors and health status under different time delays, and provide a basis for identifying the optimal time delay. According to the time delay-similarity matrix, the time delay parameters are extracted, and the time delay for each environmental feature to have the most significant impact on health can be determined. For example, the PM2.5 concentration has the greatest impact on respiratory health after a delay of 12 hours.

[0118] In an embodiment of the present invention, a reasonable time delay range is set according to professional knowledge or experience, such as 0 to 24 hours, indicating the maximum time delay considering the impact of environmental factors on health. According to the set time delay range, a time delay sequence is generated, for example, in units of hours, to generate an integer sequence containing 0 to 24. According to the time delay sequence, the environmental feature time series is shifted to construct a delayed environmental feature sequence set. For example, the PM2.5 concentration time series is shifted backward by 0 to 24 hours respectively to generate 25 PM2.5 concentration time series to form a delayed environmental feature sequence set of PM2.5 concentration. For each time series in each delayed environmental feature sequence set, and the individual health trajectory data of the patient obtained in step S28, the similarity between them is calculated respectively, for example, using a dynamic time warping (DTW) algorithm. The calculated similarity values ​​are arranged in the order of time delay and environmental features to construct a time delay-similarity matrix. For each environmental feature in the time delay-similarity matrix, the maximum similarity value in its corresponding column and the time delay corresponding to the value are found as the time delay parameter of the environmental feature. For example, if the PM2.5 concentration has the highest similarity with the patient's individual health trajectory data when delayed by 12 hours, 12 hours is used as the time delay parameter for the PM2.5 concentration.

[0119] Preferably, step S4 comprises the following steps:

[0120] Step S41: integrating the patient's multi-scale time series feature data, the patient's individual health trajectory data, and the patient's environmental sensitivity spectrum to obtain the patient's multi-dimensional data;

[0121] Step S42: performing variable selection on the multi-dimensional data of the patient to obtain a candidate variable set; performing Bayesian network structure learning based on the candidate variable set to obtain an initial network structure;

[0122] Step S43: using a preset medical expert system to integrate expert knowledge on the initial network structure to obtain a modified network structure; performing parameter learning on the modified network structure to obtain a parameterized network model;

[0123] Step S44: constructing a dynamic Bayesian network model according to the parameterized network model to obtain a dynamic Bayesian network model; extracting individualized parameters of the patient from the multi-dimensional data of the patient to obtain individualized parameters of the patient;

[0124] Step S45: Generate a personalized dynamic Bayesian network based on the dynamic Bayesian network model and the patient's individualized parameters to obtain a personalized dynamic Bayesian network.

[0125] The present invention integrates the multi-scale time series feature data, individual health trajectory data and environmental sensitivity spectrum of patients to construct multi-dimensional data of patients, which can integrate various information such as the patient's physiological state, health trajectory and environmental sensitivity, and provide a data basis for the subsequent construction of a more comprehensive and accurate prediction model. Through variable selection, the key variables related to the prediction of acute respiratory infection risk are screened out, and the initial network structure is constructed using the Bayesian network structure learning algorithm, which can automatically learn the dependencies between variables from the data, and preliminarily construct a network model that can explain the mechanism of acute respiratory infection. The initial network structure is corrected by using a medical expert system, and parameter learning is performed, which can integrate expert knowledge into the model, improve the interpretability and credibility of the model, and at the same time, learn network parameters according to the data, so that the model is more in line with the actual situation and improves the accuracy of the prediction. The parameterized network model is expanded to a dynamic Bayesian network model, and individualized parameters are extracted according to the individual data of the patient, which can capture the dynamic changes of the patient's health status, and consider individual differences, so that the model is more personalized and the prediction results are more accurate. A personalized dynamic Bayesian network is generated based on the dynamic Bayesian network model and the patient's individualized parameters, which can customize an exclusive prediction model for each patient, more accurately predict the individual's risk of acute respiratory infection in the future, and achieve personalized risk assessment and early warning.

[0126] In an embodiment of the present invention, the multi-scale time series feature data of the patient obtained in step S15, the individual health trajectory data of the patient obtained in step S28, and the environmental sensitivity spectrum of the patient obtained in step S37 are integrated to construct a unified data table containing multi-dimensional information of the patient, which is convenient for subsequent model training and prediction. The multi-dimensional data of the patient integrated in step S41 is subjected to variable selection to screen out key variables related to the risk prediction of acute respiratory infection, such as certain specific physiological indicators, time window characteristics, environmental factors, etc., to form a candidate variable set. Then, the initial Bayesian network structure is constructed based on the conditional independence relationship between the variables in the candidate variable set using the Bayesian network structure learning algorithm. The initial network structure is modified using a pre-set medical expert system, such as a knowledge map or rule base containing medical knowledge related to acute respiratory infection. For example, network nodes or edges can be added or deleted according to expert knowledge to adjust the network structure to make it more consistent with medical logic and professional knowledge. The modified network structure will be used as the final Bayesian network structure. Then, the conditional probability table of each node in the network structure is learned based on the multi-dimensional data of the patient using a Bayesian network parameter learning algorithm, such as maximum likelihood estimation or Bayesian estimation, to obtain a parameterized Bayesian network model. In order to capture the differences between individual patients, the parameterized Bayesian network model is extended to a dynamic Bayesian network model. For example, some nodes in the network can be set as dynamic nodes that change over time, and their state transition probabilities are defined so that they can simulate the dynamic changes in the patient's health status. Then, based on the multi-dimensional data of each patient, the Bayesian network inference algorithm, such as the confidence propagation algorithm or the importance sampling algorithm, is used to infer the patient's individualized network parameters, such as the initial state probability distribution of each node, the state transition probability, etc. A personalized dynamic Bayesian network is constructed based on the dynamic Bayesian network model and the patient's individualized parameters. Each patient will have a network model customized according to their own data to more accurately predict their risk of acute respiratory infections in the future.

[0127] Preferably, it is characterized in that step S5 comprises the following steps:

[0128] Step S51: collecting real-time data of patients with acute respiratory tract infection to obtain real-time data of the patients; inputting the real-time data of the patients into a personalized dynamic Bayesian network, updating the dynamic Bayesian network, and obtaining a dynamic Bayesian update network;

[0129] Step S52: predicting the risk of acute respiratory tract infection in the future according to the dynamic Bayesian update network to obtain the probability data of the risk of acute respiratory tract infection in the future;

[0130] Step S53: using a preset risk level threshold to classify the future acute respiratory infection risk probability data into risk levels to obtain risk level data;

[0131] Step S54: setting warning rules according to the risk level data and the patient's real-time data to obtain warning rule data; pushing personalized acute respiratory infection warnings according to the warning rule data to obtain personalized acute respiratory infection warning data.

[0132] The present invention can timely capture the latest changes in the patient's health status by collecting patient data in real time and updating the personalized dynamic Bayesian network, and adjust the model parameters according to the new information, so that the model can dynamically adapt to the individual situation of the patient, and improve the accuracy and real-time performance of the prediction. Using the dynamically updated Bayesian network to predict the risk probability of future acute respiratory infection can provide a more timely and accurate risk assessment, and provide a more reliable decision-making basis for patients and doctors. According to the preset risk level threshold, the risk probability is graded, and the complex probability information can be converted into a more intuitive and easier to understand risk level, which is convenient for patients and doctors to quickly judge the degree of risk. According to the risk level and the patient's real-time data, personalized warning rules are set, and warnings are pushed. According to the patient's individual situation and risk level, more accurate and personalized warning information can be provided, reminding patients to take preventive measures in time to reduce the risk of acute respiratory infection.

[0133] In an embodiment of the present invention, the physiological index data, environmental data, behavioral data, etc. of the patient with acute respiratory infection are collected in real time by wearable devices to obtain the patient's real-time data. The patient's real-time data is input into a personalized dynamic Bayesian network, and the state probability distribution and parameters of the network nodes are updated according to the new data using the Bayesian network online learning algorithm to obtain the dynamically updated Bayesian network. According to the dynamically updated Bayesian network in step S51, the probability of the patient developing acute respiratory infection in the future is calculated using a Bayesian network reasoning algorithm, such as a confidence propagation algorithm, to obtain the future acute respiratory infection risk probability data. According to the pre-set risk level threshold, such as low risk, medium risk, and high risk, the future acute respiratory infection risk probability data is graded to obtain the patient's risk level data. According to the risk level data and the patient's real-time data, such as the current physiological indicators, the degree of environmental pollution, etc., a personalized warning rule is set. For example, for high-risk patients, a stricter warning rule can be set, such as when the patient's body temperature exceeds 37.5 degrees, a high-risk warning message is immediately issued. Finally, according to the warning rules, personalized acute respiratory infection warning information is pushed to patients, for example, through mobile phone text messages, application notifications, etc., to remind patients to take preventive measures in time.

[0134] Preferably, the present invention also provides a model training system for acute respiratory infection prediction, which is used to execute the model training method for acute respiratory infection prediction as described above, and the model training system for acute respiratory infection prediction comprises:

[0135] The multi-scale time series construction module is used to collect the original data of the target patient and divide the time window to obtain the time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain the important biorhythm time window feature data; arrange the important biorhythm time window feature data in time order to obtain the multi-scale time series feature data of the patient;

[0136] The individual health trajectory modeling module is used to construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain cluster data of the patient's health trajectory; generate individual health trajectories of the patient based on the cluster data of the patient's health trajectory and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient;

[0137] The environmental sensitivity analysis module is used to obtain the patient area environmental data of the target patient through the environmental detection station to obtain the patient area environmental data; perform time dimension alignment according to the patient area environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain a patient environmental sensitivity spectrum;

[0138] A dynamic Bayesian network construction module is used to construct a dynamic Bayesian network model based on the patient's multi-scale time series feature data, the patient's individual health trajectory data, and the patient's environmental sensitivity spectrum to obtain a dynamic Bayesian network model; and to generate a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network;

[0139] The real-time risk prediction and early warning module is used to collect real-time data of patients with acute respiratory infections to obtain real-time data of patients; input the real-time data of patients into a personalized dynamic Bayesian network, update the dynamic Bayesian network, and obtain a dynamic Bayesian update network; predict future acute respiratory infection risks based on the dynamic Bayesian update network and push personalized acute respiratory infection early warnings to obtain personalized acute respiratory infection early warning data.

[0140] Preferably, a computer-readable storage medium stores a computer program, which, when executed, implements the model training method for acute respiratory infection prediction as described above.

[0141] The present invention collects, processes and integrates the electronic medical record data and physiological index data of patients, and extracts biorhythm features to construct multi-scale time series feature data of patients containing rich information, providing a high-quality data basis for subsequent model training and prediction. The VAE model is used to learn and reduce the multi-scale time series feature data of patients, and through trajectory clustering analysis and future trajectory prediction, personalized health trajectory data that can reflect the law of changes in the individual health status of patients is generated, providing more accurate individualized information for subsequent risk prediction. By analyzing the correlation between environmental data and individual health trajectory data of patients, environmental factors that have a significant impact on the health of patients are identified, and a personalized environmental sensitivity spectrum that can reflect the sensitivity of patients to different environmental factors is constructed, providing important environmental information for a comprehensive assessment of the risk of acute respiratory infections in patients. By integrating multi-scale time series feature data of patients, individual health trajectory data and environmental sensitivity spectrum, a personalized dynamic Bayesian network model that can capture individual differences, dynamic changes and environmental influences of patients is constructed using a Bayesian network, providing an effective tool for more accurate prediction of the risk of acute respiratory infections. By collecting patient data in real time and updating the personalized dynamic Bayesian network model, real-time monitoring, dynamic evaluation and personalized early warning of the risk of acute respiratory infection in patients are realized, providing more timely, more accurate and more personalized services for the health management of patients, which helps to reduce the risk of disease. Therefore, the present invention provides a model training method for acute respiratory infection prediction. In order to solve the deficiencies of traditional model training methods for acute respiratory infection prediction in depicting the dynamic changes and individual differences of patients, a series of designs such as time dimension modeling, multi-scale feature fusion, individualized representation learning, knowledge-driven optimization, and dynamic update early warning are used to capture the dynamic changes and individual differences of patients. Comprehensive consideration and modeling are carried out, which makes up for the shortcomings of traditional methods, effectively solves the problem that the existing model does not consider the individual differences and dynamic changes, and realizes more accurate and personalized prediction of acute respiratory infection risks, providing a more effective tool for disease prevention.

[0142] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

Claims

1. A model training method for predicting acute respiratory infection, characterized in that: The following steps are involved: Step S1: collect the original data of the target patient and divide it into time windows to obtain time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data; arrange the important biorhythm time window feature data in time sequence to obtain the patient's multi-scale time series feature data; Step S2: construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain clustering data of health trajectories of the patient; generate individual health trajectories of the patient based on the clustering data of health trajectories of the patient and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient; Step S3: Acquire the target patient's regional environmental data through the environmental monitoring station to obtain the patient's regional environmental data; perform time dimension alignment according to the patient's regional environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain the patient's environmental sensitivity spectrum; Step S4: constructing a dynamic Bayesian network model based on the multi-scale time series feature data of the patient, the individual health trajectory data of the patient, and the environmental sensitivity spectrum of the patient to obtain a dynamic Bayesian network model; generating a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network; Step S5: collecting real-time data of patients with acute respiratory tract infection to obtain real-time data of the patients; inputting the real-time data of the patients into a personalized dynamic Bayesian network, updating the dynamic Bayesian network, and obtaining a dynamic Bayesian update network; Based on the dynamic Bayesian update network, the risk of future acute respiratory infection is predicted and personalized acute respiratory infection warning is pushed to obtain personalized acute respiratory infection warning data.

2. The model training method for predicting acute respiratory tract infection according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining electronic medical record data of the patient through the medical system; obtaining physiological index data of the patient through the wearable device; performing data preprocessing on the electronic medical record data of the patient and the physiological index data of the patient to obtain preprocessed patient data; dividing the preprocessed patient data into time windows to obtain time window data; Step S12: extracting biorhythm features according to the time window data to obtain biorhythm feature data; performing window feature screening on the biorhythm feature data to obtain important biorhythm time window feature data; Step S13: performing multi-scale feature fusion on the important biological rhythm time window feature data to obtain multi-scale feature data of the patient; Step S14: performing time dimension feature enhancement on the multi-scale feature data of the patient to obtain multi-scale feature enhanced data of the patient; Step S15: Arrange the patient's multi-scale feature enhancement data in time sequence to obtain the patient's multi-scale time series feature data.

3. The model training method for acute respiratory infection prediction according to claim 2, characterized in that: Step S12 includes the following steps: Step S121: extracting physiological indicators from the time window data to obtain a physiological indicator sequence; Step S122: performing periodic analysis on the physiological index sequence to obtain significant periodic component data; determining the main period length according to the significant periodic component data to obtain the main period data; Step S123: extracting biorhythm features from the physiological indicator sequence according to the main cycle data to obtain biorhythm feature data; Step S124: constructing biorhythm time window features for the biorhythm feature data to obtain biorhythm time window feature data; Step S125: Filter the biological rhythm time window feature data by window feature to obtain important biological rhythm time window feature data.

4. The model training method for predicting acute respiratory tract infection according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: dividing the patient's multi-scale time series feature data into data sets to obtain a training set and a verification set; performing sliding window data augmentation and random noise addition on the training set and the verification set to obtain an enhanced training set and an enhanced verification set; Step S22: constructing a VAE model according to the multi-scale time series feature data of the patient to obtain a VAE model framework; Step S23: Using the enhanced training set to perform model training on the VAE model framework and using the enhanced validation set to perform model evaluation to obtain a VAE model; Step S24: Using the VAE model to extract latent variables from the patient's multi-scale time series feature data to obtain the patient's latent variable sequence; Step S25: Performing latent variable space visualization on the patient's latent variable sequence to obtain a patient health status distribution diagram; performing time smoothing processing on the patient's latent variable sequence to obtain a patient's latent variable smoothed sequence; Step S26: performing trajectory similarity calculation on the patient's latent variable smoothing sequence according to the patient's health status distribution diagram to obtain the patient's health trajectory similarity matrix; performing trajectory clustering analysis on the patient's health trajectory similarity matrix to obtain the patient's health trajectory clustering data; Step S27: predicting the future trajectory according to the patient's health trajectory clustering data and the patient's latent variable smoothing sequence to obtain the patient's future health trajectory data; Step S28: Generate the individual health trajectory of the patient according to the smoothed sequence of the patient's latent variables and the patient's future health trajectory data to obtain the individual health trajectory data of the patient.

5. The model training method for predicting acute respiratory tract infection according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: acquiring the patient area environmental data of the target patient through the environmental detection station to obtain the patient area environmental data; performing time alignment processing on the patient area environmental data according to the patient's individual health trajectory data to obtain time-aligned environmental data; Step S32: extracting infection risk-related environmental features from the time-aligned environmental data to obtain an environmental feature time series; performing multi-scale decomposition on the environmental feature time series to obtain a multi-scale environmental feature time series; Step S33: aligning the multi-scale environmental feature time series and the patient's individual health trajectory data in time dimension to obtain health and environment alignment data; Step S34: constructing a time delay-similarity matrix according to the time-aligned environment data and the individual health trajectory data of the patient to obtain a time delay-similarity matrix; extracting a time delay parameter according to the time delay-similarity matrix to obtain a time delay parameter; Step S35: constructing a health-environment similarity matrix according to the time delay parameter and the time delay-similarity matrix to obtain a health-environment similarity matrix; Step S36: Calculate the sensitivity score of the health and environment alignment data according to the health-environment similarity matrix to obtain environmental factor sensitivity score data; perform sensitivity feature screening according to the environmental factor sensitivity score data to obtain key environmental factor data; Step S37: construct a patient environmental sensitivity spectrum based on the key environmental factor data and the environmental factor sensitivity score data to obtain the patient environmental sensitivity spectrum.

6. The model training method for predicting acute respiratory tract infection according to claim 5, characterized in that: Step S34 includes the following steps: Step S341: defining a time delay range according to the time alignment environment data to obtain time delay range data; generating a time delay sequence according to the time delay range data to obtain a time delay sequence; Step S342: constructing a delayed environment feature sequence set for the environment feature time series according to the time delay sequence to obtain a delayed environment feature sequence set; Step S343: Calculate the time delay similarity of the delay environment feature sequence set and the patient's individual health trajectory data to obtain a time delay similarity value; construct a time delay-similarity matrix according to the time delay similarity value and the delay environment feature sequence set to obtain a time delay-similarity matrix; Step S344: extracting time delay parameters from the time delay-similarity matrix according to the delay environment feature sequence set to obtain time delay parameters.

7. The model training method for predicting acute respiratory tract infection according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: integrating the patient's multi-scale time series feature data, the patient's individual health trajectory data, and the patient's environmental sensitivity spectrum to obtain the patient's multi-dimensional data; Step S42: performing variable selection on the multi-dimensional data of the patient to obtain a candidate variable set; performing Bayesian network structure learning based on the candidate variable set to obtain an initial network structure; Step S43: using a preset medical expert system to fuse expert knowledge on the initial network structure to obtain a modified network structure; performing parameter learning on the modified network structure to obtain a parameterized network model; Step S44: constructing a dynamic Bayesian network model according to the parameterized network model to obtain a dynamic Bayesian network model; extracting individualized parameters of the patient from the multi-dimensional data of the patient to obtain individualized parameters of the patient; Step S45: Generate a personalized dynamic Bayesian network based on the dynamic Bayesian network model and the patient's individualized parameters to obtain a personalized dynamic Bayesian network.

8. The model training method for predicting acute respiratory tract infection according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: collecting real-time data of patients with acute respiratory tract infection to obtain real-time data of the patients; inputting the real-time data of the patients into a personalized dynamic Bayesian network, updating the dynamic Bayesian network, and obtaining a dynamic Bayesian update network; Step S52: predicting the risk of acute respiratory tract infection in the future according to the dynamic Bayesian update network to obtain the probability data of the risk of acute respiratory tract infection in the future; Step S53: using a preset risk level threshold to classify the future acute respiratory infection risk probability data into risk levels to obtain risk level data; Step S54: setting warning rules according to the risk level data and the patient's real-time data to obtain warning rule data; pushing personalized acute respiratory infection warnings according to the warning rule data to obtain personalized acute respiratory infection warning data.

9. A model training system for predicting acute respiratory infection, characterized in that: Used to execute the model training method for acute respiratory infection prediction according to claim 1, the model training system for acute respiratory infection prediction comprises: The multi-scale time series construction module is used to collect the original data of the target patient and divide the time window to obtain the time window data; extract the biorhythm features according to the time window data to obtain the biorhythm feature data; perform window feature screening on the biorhythm feature data to obtain the important biorhythm time window feature data; arrange the important biorhythm time window feature data in time order to obtain the multi-scale time series feature data of the patient; The individual health trajectory modeling module is used to construct a VAE model based on the multi-scale time series feature data of the patient to obtain a VAE model; use the VAE model to extract latent variables from the multi-scale time series feature data of the patient to obtain a smooth sequence of latent variables of the patient; perform trajectory clustering analysis based on the smooth sequence of latent variables of the patient to obtain cluster data of the patient's health trajectory; generate individual health trajectories of the patient based on the cluster data of the patient's health trajectory and the smooth sequence of latent variables of the patient to obtain individual health trajectory data of the patient; The environmental sensitivity analysis module is used to obtain the patient area environmental data of the target patient through the environmental detection station to obtain the patient area environmental data; perform time dimension alignment according to the patient area environmental data and the patient's individual health trajectory data to obtain health and environment alignment data; construct a time delay-similarity matrix according to the time-aligned environmental data and the patient's individual health trajectory data to obtain a time delay-similarity matrix; extract time delay parameters according to the time delay-similarity matrix to obtain time delay parameters; construct a patient environmental sensitivity spectrum according to the time delay parameters and the time delay-similarity matrix to obtain a patient environmental sensitivity spectrum; A dynamic Bayesian network construction module is used to construct a dynamic Bayesian network model based on the patient's multi-scale time series feature data, the patient's individual health trajectory data, and the patient's environmental sensitivity spectrum to obtain a dynamic Bayesian network model; and to generate a personalized dynamic Bayesian network based on the dynamic Bayesian network model to obtain a personalized dynamic Bayesian network; The real-time risk prediction and early warning module is used to collect real-time data of patients with acute respiratory infections to obtain real-time data of patients; input the real-time data of patients into a personalized dynamic Bayesian network, update the dynamic Bayesian network, and obtain a dynamic Bayesian update network; predict future acute respiratory infection risks based on the dynamic Bayesian update network and push personalized acute respiratory infection early warnings to obtain personalized acute respiratory infection early warning data.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed, the model training method for predicting acute respiratory tract infection according to any one of claims 1 to 8 is implemented.

Images