Heart failure risk prediction method and system based on multi-source heterogeneous data fusion

Abstract

The invention relates to the technical field of medical health information, in particular to a heart failure risk prediction method and system based on multi-source heterogeneous data fusion, and the prediction system comprises a data collection module, a data management module, a multi-modal feature extraction module and a dynamic risk prediction model module. An intervention strategy recommendation module; and a visualization and iterative optimization module. The prediction method is applied to the prediction system, patient health data is collected and obtained, a multi-source heterogeneous health database is constructed, physiological time sequence features, traditional Chinese medicine dialectical features and western medicine clinical features are extracted respectively to be subjected to structured coding processing, a multi-modal feature set is formed, the multi-modal feature set is divided into a training set, a verification set and a test set, and the training set, the verification set and the test set are combined. A Bayesian attention mechanism is introduced, dynamic prediction of the heart failure risk is achieved through the constructed depth time sequence model, an intervention scheme can be adjusted and formulated in the whole course of heart failure management, excessive medical treatment is reduced, meanwhile, disease progress is effectively restrained or delayed, and the life quality of a patient is improved.

Description

Technical Field

[0001] The present invention relates to the field of medical and health information technology, and particularly to a heart failure risk prediction method and system based on multi-source heterogeneous data fusion. Background Art

[0002] Heart failure is a complex clinical syndrome characterized by high incidence and high mortality. At present, the management methods for patients mainly rely on periodic offline follow-up or sporadic in-hospital monitoring, making it difficult to obtain the course evolution information in a timely and dynamic manner. At the same time, the combination of traditional Chinese and Western medicine in heart failure management has shown obvious potential value, but there is a lack of systematic and quantifiable multimodal data support and collaborative optimization strategies in clinical practice.

[0003] With the development and popularization of various channels such as wearable devices, mobile applications, and electronic medical record systems, heart failure patients can continuously collect physiological time-series data in their daily lives, such as heart rate, heart rate variability, blood oxygen saturation, activity level, sleep quality, etc., and combine subjective behavior and clinical diagnosis data such as traditional Chinese medicine syndrome differentiation information and Western medicine test results. Existing technologies often focus on single-modal or short-term monitoring, lack the full utilization of multimodal information, and lack in-depth research on uncertainty assessment, model interpretability, and dynamic intervention effects.

[0004] Therefore, there is an urgent need for a heart failure risk prediction method and system based on multi-source heterogeneous data fusion. Summary of the Invention

[0005] The purpose of the present invention is to provide a heart failure risk prediction method and system based on multi-source heterogeneous data fusion, which can take into account clinical feasibility and large-scale promotion and application value, and provide a scientific basis for improving the prognosis of heart failure and the quality of life of patients.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A heart failure risk prediction method based on multi-source heterogeneous data fusion, comprising the following steps:

[0008] S1: Respectively collect and obtain patient health data from the hospital information system, traditional Chinese medicine consultation records, mobile applications, and wearable devices as original data, and construct a multi-source heterogeneous health database;

[0009] S2: Preprocess the original data in the multi-source heterogeneous health database to obtain preprocessed data, and obtain multi-dimensional feature data therefrom to form a multi-dimensional feature data set, and store the preprocessed data, multi-dimensional feature data, and original data in layers in the multi-source heterogeneous health database;

[0010] S3: Respectively obtain physiological time-series features, traditional Chinese medicine dialectical features, and Western medicine clinical features from the multi-dimensional feature dataset, and perform structured encoding processing to obtain physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and Western medicine clinical feature vectors. Construct a multi-modal feature set and divide it into a training set, a validation set, and a test set;

[0011] S4: Construct a deep time-series model, introduce a Bayesian attention mechanism, perform weighted fusion on the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and Western medicine clinical feature vectors in the training set of the multi-modal feature set, train and optimize the deep time-series model to obtain a dynamic risk prediction model, and predict the heart failure risk of patients in the short term, medium term, or long term;

[0012] S5: Input the test set of the multi-modal feature set into the dynamic risk prediction model to obtain the heart failure risk prediction result;

[0013] S6: Based on the above heart failure risk prediction results, combined with patient-specific features, formulate personalized intervention strategies for patients;

[0014] S7: Apply the dynamic risk prediction model to dynamically monitor the safety risks of heart failure patients and dynamically adjust the intervention strategy according to the safety risks.

[0015] Preferably, the raw data in the multi-source heterogeneous database is preprocessed, including data cleaning, missing value filling, data alignment, and unique identifier management of the raw data.

[0016] Preferably, in S3, the multi-dimensional feature data is obtained by evaluating the data quality through set indicators in the preprocessed data. The multi-dimensional feature data is respectively subjected to short-term time-series slicing, medium-term time-series slicing, and long-term time-series slicing. The physiological time-series features, traditional Chinese medicine dialectical features, and Western medicine clinical features within the short-term time-series slice, medium-term time-series slice, and long-term time-series slice are respectively extracted and subjected to structured encoding processing and then stored as multi-source historical data. The physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and Western medicine clinical feature vectors are respectively extracted from the multi-source historical data by using the sliding window method to form a multi-modal feature set.

[0017] Preferably, the structured encoding processing of the physiological time-series features includes quantifying the activity level and performing Z-score standardization processing on the time-series signals to obtain physiological signal feature vectors; the structured encoding processing of the traditional Chinese medicine dialectical features includes extracting features from the traditional Chinese medicine dialectical data and syndrome type data and quantifying them into numerical values or intervals to obtain unified traditional Chinese medicine dialectical feature vectors; the structured encoding processing of the Western medicine clinical features is to standardize the Western medicine detection indicators to obtain unified Western medicine clinical feature vectors.

[0018] Preferably, S4 includes the following steps:

[0019] S41: Construct an LSTM-Attention-TCN deep time series model, use the Bayesian attention mechanism to perform weighted fusion on the importance of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and Western medicine clinical feature vectors in the training set, and combine the MC Dropout technology for uncertainty estimation to improve the confidence of the deep time series model;

[0020] S42: Input the training set and the validation set into the deep time series model for training to obtain training prediction results. Compare the training prediction results with the data in the validation set, and optimize the deep time series model according to the comparison results to obtain a dynamic risk prediction model.

[0021] Preferably, the Bayesian attention mechanism dynamically assigns weights according to the importance of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and Western medicine clinical feature vectors, and visualizes the attention weights of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and Western medicine clinical feature vectors at the output stage of the deep time series model.

[0022] Preferably, the application of the dynamic risk prediction model in S7 includes the following steps:

[0023] Heart failure patients provide real-time feedback data through a mobile application or wearable device, including medication compliance, symptom changes, and relevant physiological indicators. The feedback data of heart failure patients obtained is synchronized to the multi-modal feature set as a new test set and input into the dynamic risk prediction model, and the predicted risk results are output, and the dynamic risk prediction model is trained iteratively; and intervention strategies are formulated according to the predicted risk results; the Chinese and Western medicine treatment plans are adjusted according to the intervention strategies.

[0024] A heart failure risk prediction system based on multi-source heterogeneous data fusion, which is applied to the above-mentioned heart failure risk prediction method based on multi-source heterogeneous data fusion, includes:

[0025] A data acquisition module, configured to obtain multi-source heterogeneous health data;

[0026] A data management module, configured to process multi-source heterogeneous data and obtain multi-dimensional feature data according to the processed data;

[0027] A multi-modal feature extraction module, configured to extract physiological time series features, traditional Chinese medicine syndrome differentiation features, and Western medicine clinical features from the multi-dimensional feature dataset, and perform structured encoding processing on them to construct a multi-modal feature set;

[0028] The dynamic risk prediction model module is configured to introduce the Bayesian attention mechanism, weight and fuse the multi-modal feature set, combine the uncertainty estimation technology, train and optimize the constructed deep time series model to obtain the dynamic risk prediction model, and generate the heart failure risk prediction result;

[0029] The intervention strategy recommendation module is configured to automatically match the intervention strategy according to different heart failure syndromes and treatment stages based on the prediction result, in combination with the pre-built traditional Chinese medicine prescription library, western medicine drug specifications and acupuncture and massage program library, and generate the intervention recommendation plan;

[0030] The visualization and iterative optimization module is configured to visualize the dynamic risk prediction model through the interpretability technology, and input the real-time feedback data of the patient into the dynamic risk prediction model to obtain the predicted risk result, and dynamically adjust the intervention strategy.

[0031] Preferably, the multi-modal feature extraction module includes a physiological time series feature structured processing unit, an image analysis processing unit and a medical data analysis processing unit. The physiological time series feature structured processing unit can identify and extract various time series features. The image analysis processing unit extracts and quantifies the objective indicators of the tongue image using image processing software. The medical data analysis processing unit performs statistical analysis and standardization processing on the clinical data using statistical analysis software.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] A heart failure risk prediction method based on multi-source heterogeneous data fusion provided by the present invention constructs a multi-modal data set based on the obtained multi-source heterogeneous health data, introduces the Bayesian attention mechanism, realizes the dynamic prediction of heart failure risk by constructing a deep time series model of multi-modal data fusion, evaluates the short-term, medium-term and long-term risks of heart failure, and combines the Monte Carlo method for uncertainty quantification. By sampling the attention weight distribution of physiological time series features, traditional Chinese medicine dialectical features and western medicine clinical features, the confidence of the dynamic risk prediction model in key feature selection is evaluated, the robustness and credibility of the prediction result are improved, and the fusion representation and dynamic analysis results of the dynamic risk prediction model are comprehensively utilized, significantly improving the prediction accuracy of acute exacerbation and complications.

[0034] A heart failure risk prediction method based on multi-source heterogeneous data fusion provided by the present invention combines traditional Chinese medicine syndrome differentiation information and Western medicine indicators. The dynamic risk prediction model makes the dynamic risk prediction model more credible and interpretable in clinical applications, helps medical staff and patients understand and adopt the intervention suggestions given by the system, provides a data-driven scientific basis for the personalized design of integrated traditional Chinese and Western medicine treatment plans, and improves the diagnosis, treatment and management effects; the dynamic risk model obtains patient feedback data and performs closed-loop optimization of dynamic intervention strategies, can adjust the plan in real time throughout the entire course of heart failure management, and effectively delays the progression of the disease.

[0035] A heart failure risk prediction system based on multi-source heterogeneous data fusion provided by the present invention is applied to the above-mentioned heart failure risk prediction method based on multi-source heterogeneous data fusion, can formulate intervention plans throughout the entire course of heart failure management, reduce over-medical treatment, and improve the quality of life of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is the overall flowchart of a heart failure risk prediction method based on multi-source heterogeneous data fusion according to an embodiment of the present invention;

[0037] Figure 2 It is Figure 1 The schematic diagrams of processes S1, S2 and S3 in

[0038] Figure 3 It is Figure 1 The schematic diagrams of processes S3, S4 and S5 in

[0039] Figure 4 It is the application process schematic diagram of the dynamic risk prediction model according to an embodiment of the present invention;

[0040] Figure 5 It is the structural schematic diagram of a heart failure risk prediction system based on multi-source heterogeneous data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0042] In the description of the following invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", "left", "right", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation. The term "connection" only means the connection between devices and has no special meaning.

[0043] In addition, the technical fields and installation methods involved in the embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0044] Specific embodiments: Please refer to Figures 1-4 , a heart failure risk prediction method based on multi-source heterogeneous data fusion, comprising the following steps:

[0045] S1: Respectively collect and obtain the patient's health data from the hospital information system, traditional Chinese medicine consultation records, mobile applications, and wearable devices as the original data, and construct a multi-source heterogeneous health database.

[0046] Specifically, the multi-source heterogeneous health data includes the patient's physiological time-series data collected from wearable devices, the patient's behavior data and self-perceived symptom information collected from mobile applications, and the patient's historical information obtained from the hospital information system and traditional Chinese medicine consultation records. The collected patient health data is remotely synchronized to the multi-source heterogeneous health database as the original data.

[0047] Among them, the wearable devices include smart watches, electrocardiogram patches, or blood oxygen monitors, etc., which are used to collect the patient's physiological time-series data, such as heart rate (HR), heart rate variability (HRV), blood oxygen saturation (SPO2), activity level, and sleep status, etc.; the patient creates a personal ID on the wearable device and logs in through the personal ID to collect the physiological time-series information.

[0048] The mobile application includes the mobile APP installed on the mobile phone. Patient behavior data and self-perceived symptom information are collected through the mobile application. The patient behavior data includes daily diet records, medication compliance records, mental state, and other living habits. The daily diet records include the types, portions, calories, and nutritional components of the ingested foods. Patients are regularly prompted to record the intake of high-salt and high-fat foods. The medication compliance records include automatic reminders and manual filling of the medication time, dosage, and missed dose situations. The living habit survey includes the water intake, smoking, alcohol consumption frequency, and other special health behaviors. The mental state is regularly tested mainly through mental state questionnaires, such as the PHQ-9 and GAD-7 scales. The PHQ-9 is a short self-assessment questionnaire for evaluating depression, and the GAD-7 scale for depression is a short self-assessment questionnaire for evaluating anxiety, to assess the patient's depression and anxiety levels. The self-perceived symptom records include subjective symptoms related to heart failure, such as dyspnea, chest tightness, weight gain, etc. Patients create a personal ID on the mobile application and log in to the mobile application through the personal ID to collect relevant information.

[0049] The patient's historical information includes the clinical data corresponding to the patient's medical record number searched from the hospital information system and the traditional Chinese medicine syndrome differentiation data obtained from the traditional Chinese medicine consultation records. The clinical data includes the patient's past medical history, heart failure classification, and diagnostic information and test data of cardiac function grading, such as blood pressure, blood routine, BNP (brain natriuretic peptide) / NT-proBNP (N-terminal pro-brain natriuretic peptide), echocardiogram. The clinical data also includes the western medicine medication plan. The traditional Chinese medicine diagnostic data includes syndrome differentiation data such as inspection, auscultation and olfaction, interrogation, palpation, tongue manifestation, and pulse condition, as well as syndrome type data.

[0050] S2: Preprocess the original data in the multi-source heterogeneous health database to obtain preprocessed data, extract multi-dimensional feature data from it to form a multi-dimensional feature data set, and store the preprocessed data, multi-dimensional feature data, and original data in layers in the multi-source heterogeneous health database.

[0051] Specifically, the preprocessing of the original data includes data cleaning, missing value filling, and data alignment.

[0052] For the data cleaning of physiological time series data such as heart rate and blood oxygen collected by wearable devices, data cleaning needs to be combined with filtering. The methods of linear interpolation or time series interpolation are used to remove noise and fill short-term missing values. At the same time, by setting a reasonable threshold range, obvious outliers are identified and removed. For the data cleaning process of traditional Chinese medicine syndrome differentiation data, incomplete tongue manifestation and pulse condition records should be marked and missing data processing strategies, such as multiple imputation, exclusion, etc., are used. For self-contradictory or severely missing data, patients or doctors need to be visited for supplementation and correction, and the abnormal devices or subjects are traced and corrected.

[0053] After cleaning the above raw data and filling in the missing values, the data is aligned by synchronizing the time line. The physiological time series data, patient behavior data, self-reported symptom information, and patient history information after data cleaning and missing value filling are unified using patient identifiers and timestamps. The physiological time series data is continuously stored based on a sampling frequency of every minute or every five minutes. For the non-continuous sampling data in traditional Chinese medicine syndrome differentiation data and western medicine clinical data, a neighboring time point alignment strategy is used to achieve the alignment of multi-source data.

[0054] Among them, the patient identifier is the patient's personal ID or the medical record number in the hospital. Based on the patient's personal ID and the medical record number in the hospital, the personal ID used in the wearable device and the mobile App is identified to associate it with the patient's medical record number, ensuring that the multi-source data of the same patient in the multi-source heterogeneous database can be correctly associated.

[0055] Furthermore, from the preprocessed data, the data quality is evaluated through set indicators such as sampling coverage rate and effective recording rate, so as to obtain multi-dimensional feature data and form a multi-dimensional feature data set. The multi-dimensional feature data is used to analyze and explore its dynamic evolution law. To adapt to the time series attributes of the data in the multi-dimensional feature data set, attention needs to be paid to the alignment strategy of the sampling frequency and the timestamp to ensure that all features can accurately express the patient's physiological changes at different time periods. The raw data, preprocessed data, and multi-dimensional feature data are stored in layers and regularly backed up off-site.

[0056] The multi-dimensional feature data can be sliced into short-term time series slices, medium-term time series slices, and long-term time series slices according to research needs. The physiological time series data can be initially sliced according to the sampling timestamp of the wearable device and marked and aligned with the clinical visit time point and the traditional Chinese medicine consultation time point. The resolution of different time periods such as months, weeks, or days can be flexibly set according to research needs.

[0057] S3: Respectively obtain physiological time series features, traditional Chinese medicine syndrome differentiation features, and western medicine clinical features from the multi-dimensional feature data set, and perform structured encoding processing to obtain physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and western medicine clinical feature vectors, construct a multi-modal feature set, and divide it into a training set, a validation set, and a test set.

[0058] Specifically, from the short-term time series slices, medium-term time series slices, and long-term time series slices in the multi-dimensional feature data set, the corresponding physiological time series features, traditional Chinese medicine syndrome differentiation features, and western medicine clinical features are respectively extracted for structured processing.

[0059] Structured encoding processing method for physiological time series features. For time series signals such as heart rate, heart rate variability, and blood oxygen saturation, Z-score normalization is used to extract the time series signal feature vectors, depicting the dynamic changes of the patient's physiological state over time; for data such as activity level and sleep quality, they can be divided into different activity stages or sleep stages, such as deep sleep and light sleep, and the corresponding activity statistical feature vectors are extracted to provide the model with more dimensional information about the patient's living patterns. The fixed-dimensional features and activity statistical features are combined into a unified physiological signal feature vector;

[0060] Specifically, the Z-score normalization formula is as follows:

[0061] X = (x - μ) / σ

[0062] Where: x is the original feature value of the time series signal, X is the feature value of the time series signal after Z-score normalization; μ is the mean of the original feature values, and σ is the standard deviation of the original feature values.

[0063] Structured encoding method for traditional Chinese medicine dialectical features. For the tongue image taken by photography, objective indicators such as tongue texture, tongue coating color, tongue coating thickness, and dryness and moistness degree are extracted and quantified into numerical values or intervals to obtain the tongue image feature vector; for the pulse condition, encoding is performed using pulse rate, pulse shape fluctuation characteristics, etc. to obtain the pulse condition feature vector; based on the results of symptom inquiry, a corresponding dialectical scoring system is established for common heart failure syndromes, mapping the patient's subjective feelings and the doctor's dialectical results to specific numerical values to obtain the syndrome type feature vector. Among them, common heart failure syndromes include symptoms such as qi deficiency, yang deficiency, blood stasis, and fluid retention; the tongue image feature vector, pulse condition feature vector, and syndrome feature vector are combined into a unified traditional Chinese medicine dialectical feature vector.

[0064] The structured encoding processing method for western medicine clinical features includes normalizing or performing Z-score normalization on western medicine detection indicators such as blood pressure, BNP / NT-proBNP, cardiac ultrasound parameters, and blood routine results to obtain a unified western medicine clinical feature vector;

[0065] Furthermore, the physiological signal feature vector, traditional Chinese medicine dialectical feature vector, and western medicine clinical feature vector obtained after the above-mentioned structured processing are stored securely as multi-source historical data. The physiological signal feature vector, traditional Chinese medicine dialectical feature vector, and western medicine clinical feature vector are extracted from the multi-source historical data using the sliding window method, and a multi-modal feature set is constructed and stored.

[0066] Among them, the sliding window method is adopted to extract physiological time-series structured features, traditional Chinese medicine dialectical quantification features, and western medicine clinical structured features respectively, and the sliding window is slid according to the corresponding time dimension for three types of short-term window, medium-term window, and long-term window, corresponding to the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors in the short-term time-series slice, medium-term time-series slice, and long-term time-series slice respectively, and stored in the multi-modal feature set.

[0067] Specifically, the short-term window is set with a window length of one day to predict the risk of heart failure exacerbation in patients within 24 hours; the medium-term window is set with a window length of one week to predict the risk of readmission within one week; the long-term window is set with a window length of one month to predict the deterioration of cardiac function in long-term heart failure patients.

[0068] S4: Construct a deep time-series model, introduce the Bayesian attention mechanism, perform weighted fusion on the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors in the training set of the multi-modal feature set, train and optimize the deep time-series model to obtain a dynamic risk prediction model, and predict the short-term, medium-term, or long-term heart failure risk of patients.

[0069] The dynamic risk prediction model performs rolling prediction on the short-term, medium-term, or long-term safety risks of heart failure patients. For the prediction results of heart failure risks, a grading standard is set, and the safety level of heart failure patients within the set time period is divided into low risk, medium risk, or high risk. The risk level is automatically divided according to the probability score of the output result of the dynamic risk prediction model, providing a simple and feasible risk reminder for clinical or home intervention.

[0070] S41: Construct an LSTM-Attention-TCN deep time-series model, use the Bayesian attention mechanism to perform weighted fusion on the importance of the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors in the training set, and combine the MC Dropout technology for uncertainty estimation to improve the confidence of the deep time-series model;

[0071] Among them, the Bayesian attention mechanism replaces the deterministic weight in the attention mechanism with a probability distribution, such as a Gaussian distribution or a Dirichlet distribution, learns the posterior distribution of the attention weight through Bayesian inference, and the attention weight is adaptively adjusted over time, so that the attention weights of the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors are all probability distributions, and dynamic weighted summation is performed on the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors.

[0072] Specifically, the calculation method of dynamically weighted summation of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and western medicine clinical feature vectors is as follows: Set the physiological signal feature vector h(p), the traditional Chinese medicine syndrome differentiation feature vector h(t), and the western medicine clinical feature vector h(w), and their weighted expression is:

[0073] z = Σ(αi * h(i)), i ∈ {p, t, w};

[0074] Among them, αi is the attention weight of the i-th modality;

[0075] The parameterized expression of the attention weight of the i-th modality is:

[0076] αi = μi + σi * εi, εi ~ N(0, 1) normal distribution;

[0077] Among them, μi is the mean of the attention weight, and σi is the standard deviation of the attention weight.

[0078] The LSTM layer is used to capture the temporal dependence of physiological signals. Bayesian attention is used to dynamically adjust the weights of traditional Chinese medicine syndrome differentiation feature vectors and western medicine clinical feature vectors. The TCN module captures long-range dependencies through dilated convolutions, and residual connections are used to ensure the flow of gradients.

[0079] Furthermore, during the training process of the deep temporal model, uncertainty estimation can be performed on the deep temporal model. By using the MC Dropout technique to sample multiple times during the training stage of the deep temporal model, a relatively dispersed prediction distribution can be obtained, and the uncertainty of the output results of the deep temporal model can be quantified in the form of variance, confidence interval, etc.; for cases with low confidence, doctors should be reminded to conduct more comprehensive examinations or evaluations, and more conservative or higher safety factor intervention strategies should be given when necessary; through continuous uncertainty evaluation, the trust of medical staff in the model can be effectively improved, and at the same time, a more evidence-based and flexible heart failure intervention and management model can be provided for patients.

[0080] Among them, MC Dropout is a deep learning inference method that combines the Monte Carlo method and the Dropout regularization technique, mainly used to quantify the prediction uncertainty of neural network models. Use MC Dropout to perform uncertainty estimation on the prediction results:

[0081] Predicted mean E[y]:

[0082]

[0083] Predicted variance Var[y] (uncertainty):

[0084]

[0085] 95% confidence interval:

[0086]

[0087] Among them, M is the number of sampling times, and y(m) is the output of the m-th MC Dropout.

[0088] S42: Input the training set and the validation set into the deep time series model for training to obtain training prediction results. Compare the training prediction results with the data of the validation set, and optimize the deep time series model according to the comparison results to obtain a dynamic risk prediction model.

[0089] In the evaluation stage of the deep time series model, for multiple indicators in short-term, medium-term, and long-term prediction scenarios, such as AUC under the ROC curve, accuracy, recall rate, F1 value, etc., comprehensively measure the performance of the evaluation model according to the short-term prediction scenario, medium-term prediction scenario, and long-term prediction scenario. In addition, real-world follow-up data can be collected to verify the dynamic accuracy and applicability of the model. If there is a deviation or inaccuracy between the risk prediction result output by the deep time series model and the actual performance, corresponding data can be further subjected to reinforcement learning or active learning.

[0090] The deep time series model can adopt methods based on attention weight visualization or Grad-CAM (Gradient-weighted Class Activation Mapping) to display the region in the visualization image that has the greatest impact on the prediction result. If the dynamic risk prediction model shows inconsistencies with clinical cognition in the prediction result, targeted tracing can be carried out for this time period or feature to find the reasons for data anomalies or incorrect model inferences. The dynamic risk prediction model also combines automated dimensionality reduction techniques, such as t-SNE and UMAP, to analyze the three-dimensional feature space of traditional Chinese medicine syndromes, physiological signals, and clinical data, and confirm the most predictive feature combinations. Through this process, researchers can not only build a model with high prediction accuracy but also make it have a certain degree of clinical interpretability, providing data support for the collaborative application of traditional Chinese medicine and Western medicine in the prevention and treatment of heart failure.

[0091] S5: Input the test set in the multi-modal feature synthesis vector into the dynamic risk prediction model to obtain the heart failure risk prediction result.

[0092] Specifically, the heart failure risk prediction result includes the occurrence probability of the patient's acute heart failure attack risk, the readmission rate, and the occurrence probabilities of common complications, such as pulmonary edema, renal insufficiency, etc., and automatically divides the output heart failure risk prediction result into risk levels according to the occurrence probability.

[0093] Furthermore, the dynamic risk prediction model can also perform dynamic evolution analysis on heart failure patients. By means of time series visualization and clustering analysis methods, identify the key time nodes in the process of heart failure progression, and explore the impact of the changes of traditional Chinese and Western medicine characteristics over time on heart failure risk.

[0094] Specifically, first, the trend of the time-series output features in the risk prediction results evolving over time can be plotted as a curve or a heat map to observe the dynamic trajectory map of the significant changes in the feature values when the heart failure patients are in different clinical stages, physiological states, or traditional Chinese medicine syndrome differentiation stages, so as to show the dynamic evolution path of the patients in the three-dimensional feature space of traditional Chinese medicine syndromes, physiological signals, and clinical data;

[0095] Secondly, clustering algorithms such as K-means (K-means clustering algorithm) or DBSCAN (density clustering algorithm) are used to identify different evolution patterns of the patient population in the three dimensions of traditional Chinese medicine syndromes - physiological indicators - clinical data, so as to explore which feature combinations or time periods can best distinguish the severity of the patient's condition or the prognosis risk.

[0096] S6: Based on the above heart failure risk prediction results and combined with the patient-specific features, formulate a personalized intervention strategy for the patient.

[0097] Specifically, based on the prediction results and the patient-specific features, formulate a personalized integrated traditional Chinese and Western medicine intervention strategy. The personalized integrated traditional Chinese and Western medicine intervention strategy includes traditional Chinese medicine conditioning strategies such as traditional Chinese medicine treatment, acupuncture, meridian massage, and diet regimen, as well as Western medicine management strategies such as drug treatment, exercise prescription, and nutritional intervention.

[0098] Among them, the patient-specific features cover clinical manifestations, laboratory tests, imaging examinations, comorbidities, and dynamic changes, including the changes in BNP / NT-proBNP and troponin (ctn), the indicators of ejection fraction (EF) and ventricular remodeling markers, the symptoms of cardiorenal syndrome and arrhythmia, etc.

[0099] Both the above-mentioned traditional Chinese medicine conditioning strategy and Western medicine management strategy can design corresponding traditional Chinese medicine conditioning plans and Western medicine routine plans in advance according to the common symptoms of heart failure patients. If the patient has obvious symptoms, guidance can be provided according to the patient-specific features to ensure the coordination of the dosage and time arrangement of traditional Chinese and Western medicine.

[0100] S7: Apply the dynamic risk prediction model to dynamically monitor the safety risks of heart failure patients and dynamically adjust the intervention strategy according to the safety risks.

[0101] Specifically, heart failure patients provide real-time feedback data through a mobile application or a wearable device, such as symptom changes, medication compliance, and key physiological indicators, etc. The obtained heart failure patient feedback data is preprocessed and structurally encoded and then synchronized to the multi-modal feature set as a new test set, and input into the dynamic risk prediction model, and a new risk prediction result is output, and the dynamic risk prediction model is trained in a rolling manner; and an intervention strategy is formulated according to the new risk prediction result; the traditional Chinese and Western medicine treatment plans are adjusted according to the intervention strategy.

[0102] Among them, if it is found that the patient's risk level rises, the follow-up frequency can be encrypted, the drug dosage can be monitored more intensively, or external traditional Chinese medicine therapies can be increased, such as auricular point pressing, moxibustion, etc., to stabilize the condition; if the patient's condition remains stable or improves, the medication can be gradually reduced or some rehabilitation programs can be reduced.

[0103] According to the new risk prediction results, the trends of the triple dimensions of traditional Chinese medicine syndromes - physiological indicators - clinical data evolving over time are visualized in time series, which can further analyze the interaction between traditional Chinese medicine syndrome differentiation information and Western medicine physiological indicators; through the statistical analysis of the time series correlation between the changes in tongue and pulse conditions and key physiological indicators, where the key physiological indicators include blood oxygen and heart rate variability, the potential law of whether there is a high degree of coupling between the changes in tongue and pulse conditions and key physiological indicators in the early stage of heart failure can be explored, providing more dimensional scientific basis for the collaborative intervention of traditional Chinese and Western medicine; for example, when ecchymosis or thick tongue coating appears in traditional Chinese medicine tongue diagnosis, and the pulse is stringy and slippery, which is synchronized with the significant increase in BNP / NT-proBNP detected by Western medicine, it may indicate an early signal of the aggravation or deterioration of the condition.

[0104] Please refer to Figure 5 , a heart failure risk prediction system based on multi-source heterogeneous data fusion, which is applied to the above-mentioned heart failure risk prediction method based on multi-source heterogeneous data fusion, and includes:

[0105] The data acquisition module is respectively connected to the wearable device, the mobile application and the hospital information system, and is used to obtain multi-source heterogeneous health data from the wearable device, the mobile application and the hospital information system;

[0106] The data management module is used for cleaning, aligning, structuring, unique identification management and secure storage of multi-source heterogeneous data; the data management module adopts hierarchical permission control and desensitization technology, regularly backs up the stored data, and tracks the data access behavior through log auditing;

[0107] The multi-modal feature extraction module is used to extract physiological time series features, traditional Chinese medicine syndrome differentiation features and Western medicine clinical features from the multi-dimensional feature dataset, and perform structured coding processing and secure storage on them to construct a multi-modal feature set;

[0108] Furthermore, the multi-modal feature extraction module includes a physiological time-series feature structuring unit, an image analysis and processing unit, and a medical data analysis and processing unit. The physiological time-series feature structuring unit includes HRVAS (Heart Rate Variability Analysis Software) and Python-related libraries such as Tsfresh, or TSFEL, or Cesium. These libraries provide rich time-series feature extraction algorithms that can identify and extract various time-series features. The image analysis and processing unit uses image processing software, such as OpenCV image analysis algorithms, to extract and quantify objective indicators from tongue image pictures. The medical data analysis and processing unit uses statistical analysis software such as SPSS and SAS to perform statistical analysis and standardization processing on clinical data, etc.

[0109] The dynamic risk prediction model module is used to construct an LSTM-Attention-TCN deep time-series model, introduce a Bayesian attention mechanism, perform weighted fusion on the importance of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and western medicine clinical feature vectors, and combine uncertainty estimation techniques to train and optimize the deep time-series model to obtain a dynamic risk prediction model, generate heart failure risk prediction results, and achieve multi-time scale prediction of heart failure risks.

[0110] The intervention strategy recommendation module pre-integrates a traditional Chinese medicine prescription library, western medicine drug specifications, and an acupuncture and massage plan library, and combines the heart failure risk prediction results to automatically match and obtain intervention recommendation plans according to different heart failure syndrome types and treatment stages, so as to balance personalization and safety.

[0111] The visualization and iterative optimization module visualizes the prediction of the dynamic risk prediction model through interpretability techniques, facilitating doctors to formulate intervention strategies for patients in combination with patient-specific features and intervention recommendation plans, and inputting real-time feedback data of patients into the dynamic risk prediction model to obtain predicted risk results and dynamically adjust intervention strategies to achieve closed-loop optimization of intervention effects.

Claims

1. A heart failure risk prediction method based on multi-source heterogeneous data fusion, characterized in that: It includes the following steps: S1: Collect and obtain patient health data from the hospital information system, traditional Chinese medicine consultation records, mobile applications, and wearable devices as raw data, and construct a multi-source heterogeneous health database; S2: Preprocess the raw data in the multi-source heterogeneous health database to obtain preprocessed data, extract multi-dimensional feature data from it to form a multi-dimensional feature dataset, and store the preprocessed data, multi-dimensional feature data, and raw data in layers in the multi-source heterogeneous health database; S3: Respectively obtain physiological time-series features, traditional Chinese medicine dialectical features, and western medicine clinical features from the multi-dimensional feature dataset, perform structured encoding processing to obtain physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors, construct a multi-modal feature set, and divide it into a training set, a validation set, and a test set; S4: Construct a deep time-series model, introduce a Bayesian attention mechanism, perform weighted fusion on the physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors in the training set of the multi-modal feature set, train and optimize the deep time-series model to obtain a dynamic risk prediction model, and predict the heart failure risk of patients in the short term, medium term, or long term; S5: Input the test set of the multi-modal feature set into the dynamic risk prediction model to obtain the heart failure risk prediction result; S6: Based on the above heart failure risk prediction result, combined with patient-specific features, formulate a personalized intervention strategy for the patient; S7: Apply the dynamic risk prediction model to dynamically monitor the safety risk of heart failure patients, and dynamically adjust the intervention strategy according to the safety risk.

2. The heart failure risk prediction method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: The preprocessing of the raw data in the multi-source heterogeneous database includes data cleaning, missing value filling, data alignment, and unique identifier management of the raw data.

3. A heart failure risk prediction method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: In S3, the multi-dimensional feature data is obtained by evaluating the data quality through set indicators in the preprocessed data. The multi-dimensional feature data is respectively sliced into short-term time series, medium-term time series, and long-term time series. The physiological time-series features, traditional Chinese medicine dialectical features, and western medicine clinical features within the short-term time series slice, medium-term time series slice, and long-term time series slice are respectively extracted and subjected to structured encoding processing, and then stored as multi-source historical data. The physiological signal feature vectors, traditional Chinese medicine dialectical feature vectors, and western medicine clinical feature vectors are respectively extracted from the multi-source historical data by using the sliding window method to form a multi-modal feature set.

4. A heart failure risk prediction method based on multi-source heterogeneous data fusion according to claim 3, characterized in that: The structured encoding processing of the physiological time-series features includes quantifying the activity level and performing Z-score standardization processing on the time-series signal to obtain the physiological signal feature vector; the structured encoding processing of the traditional Chinese medicine dialectical features includes extracting features from the traditional Chinese medicine dialectical data and syndrome type data and quantifying them into numerical values or intervals to obtain a unified traditional Chinese medicine dialectical feature vector; the structured encoding processing of the western medicine clinical features is to standardize the western medicine detection indicators to obtain a unified western medicine clinical feature vector.

5. The heart failure risk prediction method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: S4 includes the following steps: S41: Construct an LSTM-Attention-TCN deep time series model, use the Bayesian attention mechanism to perform weighted fusion on the importance of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and western medicine clinical feature vectors in the training set, and combine the MC Dropout technique for uncertainty estimation to improve the confidence of the deep time series model; S42: Input the training set and the validation set into the deep time series model for training to obtain training prediction results. Compare the training prediction results with the data of the validation set, and optimize the deep time series model according to the comparison results to obtain a dynamic risk prediction model.

6. The heart failure risk prediction method based on multi-source heterogeneous data fusion according to claim 1, wherein: The Bayesian attention mechanism dynamically assigns weights according to the importance of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and western medicine clinical feature vectors, and visually displays the attention weights of physiological signal feature vectors, traditional Chinese medicine syndrome differentiation feature vectors, and western medicine clinical feature vectors at the output stage of the deep time series model.

7. The heart failure risk prediction method based on multi-source heterogeneous data fusion according to claim 1, wherein: The application of the dynamic risk prediction model in S7 includes the following steps: Heart failure patients use a mobile application or wearable device to provide real-time feedback data, including medication compliance, symptom changes, and relevant physiological indicators. The obtained heart failure patient feedback data is synchronized to the multi-modal feature set as a new test set and input into the dynamic risk prediction model, and the predicted risk result is output, and the dynamic risk prediction model is trained iteratively; And formulate an intervention strategy according to the predicted risk result; Adjust the traditional Chinese and western medicine treatment plans according to the intervention strategy.

8. A heart failure risk prediction system based on multi-source heterogeneous data fusion, which is applied to the heart failure risk prediction method based on multi-source heterogeneous data fusion according to any one of claims 1-7 above, and is characterized in that: Including: A data acquisition module configured to obtain multi-source heterogeneous health data; A data management module configured to process multi-source heterogeneous data and obtain multi-dimensional feature data according to the processed data; A multi-modal feature extraction module configured to extract physiological time series features, traditional Chinese medicine syndrome differentiation features, and western medicine clinical features from the multi-dimensional feature dataset, perform structured coding processing on them, and construct a multi-modal feature set; A dynamic risk prediction model module configured to introduce a Bayesian attention mechanism, perform weighted fusion on the multi-modal feature set, combine uncertainty estimation techniques, train and optimize the constructed deep time series model to obtain a dynamic risk prediction model, and generate a heart failure risk prediction result; An intervention strategy recommendation module configured to automatically match intervention strategies according to different heart failure syndrome types and treatment stages based on the heart failure risk prediction result, combined with a pre-built traditional Chinese medicine prescription library, western medicine drug specifications, and acupuncture and massage plan library, and generate an intervention recommendation plan; A visualization and iterative optimization module configured to visualize the dynamic risk prediction model through interpretability techniques, and input the patient's real-time feedback data into the dynamic risk prediction model to obtain the predicted risk result and dynamically adjust the intervention strategy.

9. The heart failure risk prediction system based on multi-source heterogeneous data fusion according to claim 8, characterized in that: The multimodal feature extraction module includes a physiological time series feature structuring processing unit, an image analysis processing unit, and a medical data analysis processing unit. The physiological time series feature structuring processing unit can identify and extract various time series features. The image analysis processing unit extracts and quantifies the objective indicators of tongue image pictures using image processing software. The medical data analysis processing unit performs statistical analysis and standardization processing on clinical data using statistical analysis software.

Images