Heart failure risk grading early warning method and system based on heterogeneous data fusion

Abstract

The invention discloses a heart failure risk grading early warning method and system based on heterogeneous data fusion. The method comprises the following steps: synchronously collecting three types of heterogeneous data of high-frequency physiology, low-frequency subjective reports and event-driven behaviors of a patient; a dynamic weighted fusion architecture of a long short-term memory network LSTM and a gradient boosting decision tree GBDT is adopted for processing, and dynamic fusion is carried out through an attention mechanism; clinical key events such as unplanned re-admission and the like are used as feedback signals, and closed-loop optimization is performed on the model through a specific event-triggered local online incremental learning mechanism. In a multi-center prospective clinical research involving 1258 patients, the sensitivity of unplanned re-admission risk prediction within 30 days is improved to 0.92 + / -0.04 from 0.75 of a conventional optimization method, the AUC (area under a working characteristic curve) of a subject reaches 0.95, and an innovative technical scheme is provided for realizing accurate and dynamic extra-hospital heart failure risk management.

Description

Technical Field

[0001] The present invention relates to the fields of medical information technology and artificial intelligence, and in particular to a heart failure risk grading and early warning method and system based on heterogeneous data fusion for chronic disease management. Background Art

[0002] Heart failure is the terminal stage of various heart diseases, characterized by high morbidity, high rehospitalization rate and high mortality rate, which brings a heavy medical burden to society. Currently, the outpatient management of heart failure patients after discharge generally follows the clinical practice consensus established by my country's heart failure diagnosis and treatment guidelines, that is, the condition assessment and drug dosage adjustment of heart failure patients must be completed in regular outpatient follow-up led by doctors.

[0003] However, those skilled in the art generally regard the principle of regular follow-up as the axiom for achieving the optimal balance of outpatient management. However, strict adherence to this principle directly leads to an irreconcilable cost-time-safety paradox: First, this principle leads to a contradiction between cost and timeliness. In order to achieve the ideal drug titration frequency recommended by the guidelines, patients need to pay huge time and financial costs to frequently travel to the hospital, and at the same time bring overwhelming outpatient pressure to the medical system; in actual operation, the follow-up interval is often forced to be extended to several months, but this directly sacrifices the timeliness of treatment, causing patients to miss the best window for optimizing medication during the stable period of disease. Clinical data show that more than 25% of patients are readmitted to the hospital due to worsening of their condition within 6 months after discharge, which is closely related to untimely follow-up.

[0004] Secondly, this principle brings about a safety paradox. This principle aims to ensure patient safety through direct medical consultation by doctors, but its essence is a static snapshot-style assessment model that completely ignores the dynamic evolution of the condition that may occur between two follow-up visits. For example, the patient may have experienced early signs of deterioration such as weight gain for several consecutive days and paroxysmal dyspnea at night, but these key warning information is ignored because it cannot be transmitted to the doctor in real time. This makes a management model designed to ensure safety become a major safety hazard that leads to acute decompensation events in patients due to the lag in information transmission.

[0005] After searching, it was found that the method recommended in a prior art patent with publication number CN117877681A on the China Patent Network, entitled "A method and device for monitoring and early warning of daily dietary data of patients with heart failure", itself constitutes a reverse teaching, that is, it overemphasizes offline and regular formal safety, which in turn hinders technical personnel in this field from exploring the use of real-time and dynamic data for out-of-hospital risk management, thus failing to solve the fundamental problem of out-of-hospital management.

[0006] Therefore, how to break the above paradox and build a new management model that can conduct real-time, dynamic and accurate out-of-hospital risk classification and early warning is a technical problem that needs to be solved urgently in this field. Summary of the Invention

[0007] The purpose of the present invention is to provide a heart failure risk classification and early warning method and system based on heterogeneous data fusion to solve the cost-timeliness-safety paradox problem existing in the background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions: A heart failure risk classification and early warning method based on heterogeneous data fusion, the method comprising: Step S101: synchronously collect three types of heterogeneous data for characterizing the patient's out-of-hospital status, wherein the heterogeneous data include high-frequency physiological data for characterizing the patient's objective physiological state, low-frequency subjective report data for characterizing the patient's subjective feelings and quality of life, and event-driven behavioral data for characterizing the patient's compliance with the treatment plan; Step S102: Processing the heterogeneous data using a dynamic weighted fusion architecture of a long short-term memory network (LSTM) and a gradient boosted decision tree (GBDT), wherein the LSTM model is configured to process the time series features of the high-frequency physiological data, and the GBDT model is configured to process the structured features of the low-frequency subjective report data and the event-driven behavioral data; Step S103: Dynamically adjust the output weights of the LSTM model and the GBDT model according to the real-time characteristics of the input data through a decision fusion layer based on the attention mechanism, and generate a unified risk level; Step S104: Based on predefined clinical key events as feedback signals, the decision fusion layer parameters of the dynamic weighted fusion architecture are continuously optimized through a local online incremental learning mechanism triggered by events.

[0009] Through this solution, heterogeneous data covering the three dimensions of physiology, psychology, and behavior are collected simultaneously, and a comprehensive perception system of the patient's out-of-hospital status is constructed, overcoming the defect of traditional methods that rely only on single-dimensional data and lead to one-sided evaluation.

[0010] By adopting a dynamic weighted fusion architecture of LSTM and GBDT, LSTM is used to specifically process the time dependency of high-frequency physiological data, and GBDT is used to process the structured features of low-frequency subjective reports and behavioral data, achieving optimal modeling of different data modalities and avoiding performance loss caused by feature mismatch when a single model processes heterogeneous data. Through the decision fusion layer based on the attention mechanism, the weights of each model output are dynamically adjusted, enabling the system to intelligently determine which data source is more decisive in a specific situation, thereby improving the accuracy and robustness of risk assessment.

[0011] Finally, an event-triggered local online incremental learning mechanism based on feedback from key clinical events was introduced, and medical-specific parameter adjustment rules were defined to build a closed-loop self-optimization system. This system enables the model to continuously learn and evolve from real-world clinical results, completely resolving the fundamental problem that the static snapshot evaluation model of existing technologies cannot adapt to the dynamic changes in patients' conditions.

[0012] Preferably, the high-frequency physiological data include heart rate, pulmonary edema index obtained by analyzing heart sound characteristics, blood pressure, blood oxygen saturation and weight; through this solution, key physiological indicators such as pulmonary edema index are included in the monitoring, which can effectively capture early signs of deterioration of the condition of heart failure patients caused by fluid retention, thereby improving the sensitivity of the early warning.

[0013] Preferably, the low-frequency subjectively reported data are response data from the Kansas City Cardiomyopathy Questionnaire (KCCQ-12). This approach uses the internationally recognized and U.S. Food and Drug Administration-certified KCCQ-12 questionnaire as a data source, ensuring the standardization, validity, and reliability of the subjectively reported data and providing high-quality input for the model, thereby enhancing the ability to quantitatively assess patients' quality of life and symptom burden.

[0014] Preferably, the attention-based decision fusion layer is configured to dynamically adjust the output weights of the LSTM model and the GBDT model based on the time series volatility of the high-frequency physiological data or the recent trend of changes in the low-frequency subjective report data. This solution provides the decision fusion layer with the ability to intelligently arbitrate. For example, if a patient's weight is monitored to fluctuate dramatically in the short term, the system will automatically increase the weight of the LSTM model, prioritizing the deterioration of objective physiological indicators. If a patient's KCCQ-12 score continues to decline, the weight of the GBDT model will be increased to ensure timely capture of changes in subjective perception.

[0015] Preferably, the local online incremental learning mechanism specifically includes: when the clinical key event is unplanned readmission, only the decision fusion layer weight is updated, and the physiological data loss item in the last 7 days is multiplied by a magnification factor of 1.5; when the clinical key event is successful drug titration, only the decision fusion layer weight is updated, and the feature importance threshold of the behavioral data in GBDT is increased by 40%; through this solution, the specific technical rules of model optimization are clarified, the reproducibility and stability of the online incremental learning process are ensured, and the simple application of conventional algorithms is avoided.

[0016] Preferably, the key clinical events include unplanned readmission events as negative feedback and successful drug titration events as positive feedback. Through this solution, the two most important types of outcome events in clinical practice are used as feedback signals of the model, so that the optimization direction of the model is highly consistent with the clinical goal. The use of the hard endpoint event of unplanned readmission as negative feedback can strongly drive the model to learn and identify high-risk patterns. At the same time, the innovative introduction of successful drug titration as positive feedback enables the model to learn and identify patterns of stable or improved condition, thereby achieving comprehensive optimization of the combination of rewards and punishments.

[0017] Preferably, the drug titration success event is defined as: within 14 days after the drug dose adjustment, the patient's N-terminal B-type natriuretic peptide precursor NT-proBNP level decreases by more than 30%, and their New York Heart Association NYHA heart function class improves by at least 1 level, and their 6-minute walk distance increases by at least 50 meters, and no serious adverse reactions are reported; through this scheme, an objective, quantifiable, and multi-dimensional clinical definition of a drug titration success event is provided, avoiding the ambiguity caused by fluctuations in a single biochemical indicator and ensuring the accuracy and effectiveness of the positive feedback signal.

[0018] Preferably, the risk level output by the method is only used to assist decision-making in out-of-hospital management and does not directly constitute a diagnosis or treatment plan for heart failure. Through this solution, the legal and ethical boundaries of the present invention are clarified to ensure that it serves as a technical tool to assist doctors in decision-making rather than replacing doctors' diagnostic rights. This complies with current regulatory requirements for artificial intelligence medical devices and reduces compliance risks in applications.

[0019] The present invention provides a heart failure risk grading and early warning system based on heterogeneous data fusion, the system comprising: a data acquisition module for performing the operation described in step S101 of the aforementioned method; a data processing module for performing the operation described in step S102 of the aforementioned method; a decision fusion module for performing the operation described in step S103 of the aforementioned method; and a model optimization module for performing the operation described in step S104 of the aforementioned method, the model optimization module comprising an event trigger, a parameter freezing unit, and a weight amplification and quantization unit.

[0020] The present invention has the following significant beneficial effects: The present invention produces unexpected technical effects through a specific algorithm coupling structure and closed-loop optimization mechanism. In a multicenter prospective clinical study involving 1,258 patients, the present invention's solution was compared with two control solutions and an ablation experiment solution. The results showed that the present invention's solution achieved a sensitivity index of 0.92±0.04 in predicting the risk of 30-day readmission, compared with 0.75 of the control solution representing the conventional optimization idea, achieving a significant performance leap of more than 22%, demonstrating the superiority of the present invention's heterogeneous model coupling and dynamic fusion architecture, which is by no means a simple technical superposition; in addition, compared with 0.82 of the ablation experiment solution with the online learning module removed, the performance of the complete solution of the present invention was improved by more than 12%, which irrefutably proves that the closed-loop feedback mechanism of the present invention is the key technical feature that produces unexpected technical effects and solves the fundamental defect of the static snapshot of the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the process of the heart failure risk classification and early warning method according to the present invention; Figure 2 Schematic diagram of the system hardware architecture and data flow of the present invention; Figure 3 2 is a comparison diagram of the receiver operating characteristic (ROC) curves of the experimental group and the control group of the present invention. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0023] Example 1 Reference Figure 2 The present invention provides a hardware architecture and data flow of a heart failure risk grading warning system, wherein the system architecture includes a perception layer 10, a transmission layer 20, a service layer 30 and an application layer 40.

[0024] The perception layer 10 is the source of data collection and is mainly composed of terminal devices worn or used by patients. Specifically, it may include multi-parameter vital signs monitoring equipment 11, which can continuously monitor and record high-frequency physiological data such as heart rate, blood pressure, and blood oxygen saturation; in addition, it also includes an application 12 installed on a smartphone. Patients can use this application to regularly fill out the KCCQ-12 questionnaire and record their daily medication use, thereby collecting low-frequency subjective report data and event-driven behavioral data.

[0025] The transport layer 20 is responsible for securely and reliably transmitting the data collected by the perception layer 10 to the cloud. The multi-parameter vital signs monitoring device 11 is connected to the smartphone application 12 via Bluetooth low energy technology. The application 12 then transmits the encrypted data packet to the service layer 30 via the cellular network or wireless LAN. The data transmission process uses a standard encryption protocol to ensure the confidentiality and integrity of the data during transmission.

[0026] The service layer 30 is the core of the present invention and is deployed on a cloud server. The service layer 30 includes a data acquisition module 301, a data processing module 302, a decision fusion module 303, a model optimization module 304 and a data compliance and privacy protection module 305.

[0027] The data acquisition module 301 is responsible for receiving and parsing data from the transmission layer 20. The data processing module 302 has built-in LSTM models and GBDT models to perform feature extraction and preliminary modeling on the pre-processed high-frequency physiological data and low-frequency structured data respectively; the decision fusion module 303 adopts the attention mechanism to dynamically weighted fuse the output of the data processing module 302 to generate the final risk level; the model optimization module 304 executes an online incremental learning algorithm based on the clinical key events fed back from the application layer 40, and continuously optimizes the model parameters in the decision fusion module 303 and the data processing module 302; the data compliance and privacy protection module 305 performs desensitization and pseudonymization throughout the data life cycle, and can support the federated learning framework to conduct multi-center model training while protecting patient privacy.

[0028] The application layer 40 is the interactive interface between the system and the user. For medical staff, they can view the patient's risk level, historical data trends and health recommendations through the medical workstation 41. When the risk level output by the system is high risk, a high-risk alarm can be automatically pushed to the medical workstation 41, and an instruction for a medical return visit within one hour can be triggered; for patients, personalized health guidance and medication reminders can be received through the smartphone application 12.

[0029] Example 2 Reference Figure 1 The present invention provides a heart failure risk classification and early warning method process, which specifically includes the following steps: S101. Synchronous collection of heterogeneous data. The system synchronously collects three types of heterogeneous data through the perception layer 10 described in Example 1. High-frequency physiological data includes heart rate, pulmonary edema index obtained through heart sound feature analysis, blood pressure, blood oxygen saturation, and weight measured at a fixed time every morning. The pulmonary edema index is calculated by first extracting the energy proportion of the third heart sound S3 component in the heart sound signal through wavelet transform technology, and combining it with the respiratory rate of change in the same time period to obtain the pulmonary edema index through a weighted formula. The calculation formula is: pulmonary edema index = 0.7 × S3 energy proportion + 0.3 × respiratory rate of change. Low-frequency subjective report data is the KCCQ-12 questionnaire completed by patients once a week through the application 12. Event-driven behavioral data is the medication compliance data recorded by patients daily through the application 12 or automatically reported by the smart medicine box.

[0030] S102, based on the hybrid model processing of LSTM and GBDT, the collected data is processed in the data processing module 302 of the service layer 30; for high-frequency physiological data, preprocessing is first performed, including missing value interpolation, outlier removal and smoothing filtering, and then the processed multidimensional time series is input into a pre-trained LSTM model. The LSTM network can effectively capture the long-term dependencies and time patterns in physiological signals, which is crucial for identifying the trend of heart failure from stable to slowly worsening; the LSTM model outputs a feature vector representing the patient's recent physiological state risk For low-frequency subjective report data and behavioral data, they are input into a GBDT model as structured features. GBDT is an integrated learning algorithm that is good at processing tabular data and can effectively mine the nonlinear relationships and interactions between different features. The GBDT model outputs a feature vector representing the risk of the patient's subjective and behavioral status. .

[0031] S103, decision fusion based on attention mechanism, the two feature vectors output by the data processing module 302 and It is sent to the decision fusion module 303. The core of this module is an attention mechanism layer. This mechanism calculates the relative importance of the outputs of the two models at the current moment through learnable parameters. Its weight calculation can be simplified as follows: in, and is the weight matrix for feature transformation, V is the weight vector for calculating attention scores, The function normalizes the scores into weights; the final fused feature vector It is obtained by the weighted sum of the outputs of the two models and passes through a fully connected layer and This mechanism makes the fusion process dynamic and adaptive, and can intelligently focus on more reliable information sources based on the characteristics of the input data.

[0032] S104, online incremental learning based on clinical event feedback, the model optimization module 304 is responsible for the continuous optimization of the model; when the system receives a clinical key event feedback signal from the application layer 40, for example, a patient predicted to be at medium risk has an unplanned readmission, the system will trigger an online learning; the present invention defines specific medical-specific parameter adjustment rules: when a negative feedback signal of unplanned readmission is received, the system increases the weight coefficient α related to the high-frequency physiological data of the last 7 days in the LSTM model loss function to 1.5 times to enhance the model's sensitivity to recent physiological deterioration signals; when a positive feedback signal of successful drug titration is received, the system increases the feature importance threshold β of the behavioral data in the GBDT model by 40% to strengthen the weight of the good compliance pattern. This closed-loop feedback mechanism enables the model to learn from errors and dynamically adapt to changes in patient groups and new risk patterns, thereby continuously improving the accuracy of predictions.

[0033] Example 3 To validate the technical effectiveness of the present invention, a multicenter, prospective, controlled study was conducted. A total of 1,258 discharged patients with heart failure were enrolled and randomly divided into an experimental group and a control group. Furthermore, to further analyze the contribution of the key technical features of the present invention, two additional control groups were set up for comparison. All performance indicators were evaluated on independent test sets. The experimental grouping and protocol are described below: The experimental group fully implemented the technical solution of the present invention, i.e., adopted LSTM-GBDT dynamic weighted fusion plus a closed-loop optimization architecture based on clinical event feedback.

[0034] Control group 1 completely reproduced the technical solution in a publication number CN117877681A published on the Chinese Patent Online. This solution mainly relies on setting fixed thresholds for a single or a few physiological parameters and combining it with a simplified scoring card model for risk assessment.

[0035] Control group 2 represents the direct technical combination that technical personnel in this field may adopt when facing the same problem. The three types of heterogeneous data used in the present invention are subjected to conventional feature engineering processing, and then all features are spliced ​​and input into a single GBDT model for training and prediction. This solution does not include heterogeneous model processing and dynamic fusion mechanism.

[0036] The ablation experiment group adopted the LSTM-GBDT dynamic weighted fusion architecture of the present invention, but removed the online incremental learning mechanism in the model optimization module 304. That is, the model remained static after the initial training and no longer updated its parameters based on subsequent clinical events. This group aimed to verify the key contribution of the closed-loop optimization mechanism.

[0037] After rigorous verification on the test set, the quantitative results described in the above summary of the invention were obtained. Figure 3 The area under the receiver operating characteristic (ROC) curve (AUC) of the experimental group of the present invention reached 0.95, which was significantly better than 0.68 of control group 1 and 0.84 of control group 2. These data clearly show that the present invention has achieved a nonlinear performance leap through its unique dynamic fusion of heterogeneous models and closed-loop optimization design, producing technical effects that cannot be predicted by existing technologies or their simple combinations.

[0038] The algorithm performance in this embodiment has passed the software special test of the Medical Device Testing Institute of the China Food and Drug Inspection Institute and is recorded in the artificial intelligence medical device software quality evaluation report. This system strictly abides by the relevant provisions of the Personal Information Protection Law and the Data Security Law throughout the entire process of data collection, transmission, and processing. All data are collected after the patient signs an electronic informed consent form and are pseudonymized and desensitized. When conducting multi-center data fusion analysis, privacy computing technologies such as federated learning are used to ensure that the original data does not leave the local area, thereby protecting data security and patient privacy.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A heart failure risk classification and early warning method based on heterogeneous data fusion, characterized by: include: Step S101: synchronously collecting three types of heterogeneous data for characterizing the patient's out-of-hospital status, wherein the heterogeneous data include high-frequency physiological data for characterizing the patient's objective physiological state, low-frequency subjective report data for characterizing the patient's subjective feelings and quality of life, and event-driven behavioral data for characterizing the patient's compliance with the treatment plan; Step S102: Processing the heterogeneous data using a dynamic weighted fusion architecture of a long short-term memory network (LSTM) and a gradient boosted decision tree (GBDT), wherein the LSTM model is configured to process the time series features of the high-frequency physiological data, and the GBDT model is configured to process the structured features of the low-frequency subjective report data and the event-driven behavioral data; Step S103: Dynamically adjust the output weights of the LSTM model and the GBDT model according to the real-time characteristics of the input data through a decision fusion layer based on the attention mechanism, and generate a unified risk level; Step S104: Based on predefined clinical key events as feedback signals, the decision fusion layer parameters of the dynamic weighted fusion architecture are continuously optimized through a local online incremental learning mechanism triggered by events.

2. A heart failure risk classification and early warning method based on heterogeneous data fusion according to claim 1, characterized in that: The high-frequency physiological data include heart rate, pulmonary edema index obtained by analyzing heart sound characteristics, blood pressure, blood oxygen saturation and weight; the low-frequency subjective report data are response data of the Kansas City Cardiomyopathy Questionnaire KCCQ-12.

3. The heart failure risk classification and early warning method based on heterogeneous data fusion according to claim 1 is characterized in that: The attention-based decision fusion layer is configured to dynamically adjust the output weights of the LSTM model and the GBDT model according to the time series volatility of the high-frequency physiological data or the recent change trend of the low-frequency subjective report data.

4. The heart failure risk classification and early warning method based on heterogeneous data fusion according to claim 1 is characterized in that: The clinical key events include unplanned readmission events as negative feedback and successful drug titration events as positive feedback, and the local online incremental learning mechanism specifically includes: when the clinical key event is unplanned readmission, only the decision fusion layer weights are updated, and the physiological data loss item in the last 7 days is multiplied by a magnification factor of 1.5; when the clinical key event is successful drug titration, only the decision fusion layer weights are updated, and the feature importance threshold of behavioral data in GBDT is increased by 40%.

5. The heart failure risk classification and early warning method based on heterogeneous data fusion according to claim 4 is characterized in that: The successful drug titration event was defined as: within 14 days after the drug dose adjustment, the patient's N-terminal B-type natriuretic peptide precursor NT-proBNP level decreased by more than 30%, and his New York Heart Association NYHA heart function class improved by at least 1 level, and his 6-minute walk distance increased by at least 50 meters, and no serious adverse reactions were reported.

6. The heart failure risk classification and early warning method based on heterogeneous data fusion according to claim 1 is characterized in that: The risk level output by the method is only used to assist in decision-making for out-of-hospital management and does not directly constitute a diagnosis or treatment plan for heart failure.

7. A heart failure risk grading and early warning system based on heterogeneous data fusion, characterized by: include: A data acquisition block module, configured to perform the operation described in step S101 of claim 1; A data processing module, configured to perform the operation described in step S102 of claim 1; a decision fusion module, configured to perform the operation described in step S103 of claim 1; A model optimization module is used to perform the operation described in step S104 of claim 1, wherein the model optimization module includes an event trigger, a parameter freezing unit, and a weight amplification and quantization unit.

8. A heart failure risk grading and early warning system based on heterogeneous data fusion according to claim 7, characterized in that: The data processing module, decision fusion module and model optimization module are deployed on a cloud server; the data acquisition module includes a wearable biosensor and an intelligent terminal device worn by the patient and connected to the cloud server for communication.

9. The heart failure risk grading and early warning system based on heterogeneous data fusion according to claim 7, characterized in that: Also includes: The data compliance and privacy protection module is used to desensitize data throughout the entire process of data collection, transmission, and processing, and uses a federated learning framework to protect patient data privacy.

Images