Heart failure patient remote monitoring and early warning system based on Internet of Things

Abstract

The invention discloses a heart failure patient remote monitoring and early warning system based on the Internet of Things, and the system comprises a data collection and preprocessing module which is used for collecting and preprocessing multi-source physiological and behavior data, and generating an input sample set; the feature evaluation module is used for constructing an extreme gradient boosting tree algorithm model and obtaining feature importance scores; the feature screening module is used for executing a recursive feature elimination process and outputting an optimal feature subset; the risk judgment module is used for inputting the optimal feature subset into an extreme gradient boosting tree algorithm model to generate a heart failure risk score; the early warning triggering module is used for triggering early warning and sending early warning information to the patient and the platform; the feedback recording module is used for generating feedback information; and the dynamic optimization module is used for updating the input sample set based on the feedback information, retraining the extreme gradient boosting tree algorithm model and screening the features. According to the method, the extreme gradient boosting tree and recursive feature elimination are fused, and accurate evaluation and dynamic early warning of the multi-source data of the heart failure patient are realized.

Description

Technical Field

[0001] The present invention relates to the technical fields of medical health monitoring and artificial intelligence, and particularly to a remote monitoring and warning system for heart failure patients based on the Internet of Things. Background Art

[0002] At the present stage, with the aggravation of population aging and the continuous growth of the number of heart failure patients, how to achieve effective monitoring and early warning of chronic heart failure patients has become one of the core issues in the fields of clinical medicine and home health management. Traditional heart failure risk assessment mostly relies on artificial intervention and regular examinations in hospitals, which is difficult to capture the dynamic changes of patients' physical signs in a timely manner, and has significant limitations such as data lag, slow response, and lack of continuity. With the development of Internet of Things technology, more and more wearable devices and health sensors are applied to vital sign monitoring, making it possible to collect multi-source physiological and behavioral data such as heart rate, blood pressure, weight, sleep, and activity in the home environment of patients, providing a new technical basis for the remote dynamic management of heart failure.

[0003] Some existing studies have tried to apply machine learning methods to the prediction of heart failure risk, such as establishing scoring models using algorithms such as logistic regression, support vector machine, or random forest. However, these models often face problems such as feature redundancy, poor model stability, and weak generalization ability when dealing with high-dimensional time-series data. Especially in multi-source Internet of Things monitoring data, redundant features not only can lead to model overfitting, but may also obscure key variables crucial for risk judgment, seriously affecting the accuracy and timeliness of risk prediction. In addition, existing methods generally lack an adaptive optimization mechanism for the dynamic change of model feature importance, and it is difficult to achieve automatic maintenance and continuous optimization of model performance during long-term operation. At the same time, most systems fail to establish an effective feedback learning mechanism, and cannot use historical warning results and actual processing effects for model correction and sample update, resulting in the model being difficult to adapt to the long-term changes and individual differences of patients' states.

[0004] Therefore, how to provide a remote monitoring and warning system for heart failure patients based on the Internet of Things is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] An object of the present invention is to propose a remote monitoring and warning system for heart failure patients based on the Internet of Things. The present invention makes full use of the Internet of Things multi-source health data collection technology and integrated learning modeling method, and details the whole process mechanism of feature importance evaluation, feature screening, risk scoring, multi-level warning, and dynamic feedback optimization, with advantages such as high evaluation accuracy, strong self-adaptability of feature screening, sustainable optimization of the model, and timely warning response.

[0006] The remote monitoring and warning system for heart failure patients based on the Internet of Things according to the embodiments of the present invention includes the following modules: A data acquisition and preprocessing module, which is used to acquire multi-source physiological and behavioral data and perform preprocessing to generate an input sample set; A feature evaluation module, which is used to construct an extreme gradient boosting tree algorithm model and train it to obtain feature importance scores; A feature screening module, which is used to execute a recursive feature elimination process and output an optimal feature subset; A risk discrimination module, which is used to input the optimal feature subset into the extreme gradient boosting tree algorithm model to generate a heart failure risk score; An early warning trigger module, which is used to compare the heart failure risk score with a preset risk threshold, trigger an early warning and send an early warning message to the patient and the platform; A feedback recording module, which is used to record the early warning score, early warning level and platform response result to generate feedback information; A dynamic optimization module, which is used to update the input sample set based on the feedback information and periodically retrain the extreme gradient boosting tree algorithm model and screen features.

[0007] Optionally, the modules are implemented by the following methods: S1. Collect multi-source physiological and behavioral data of heart failure patients through Internet of Things devices, preprocess the multi-source physiological and behavioral data, and construct a structured input sample set; S2. Construct an extreme gradient boosting tree algorithm model, train it through the input sample set, and obtain the feature importance scores corresponding to each input feature; S3. Based on the feature importance scores, execute a recursive feature elimination process, eliminate the feature with the lowest score in each iteration and retrain the extreme gradient boosting tree algorithm model until the set optimal feature subset scale or performance convergence condition is met, and output the optimal feature subset; S4. Based on the optimal feature subset, perform discrimination processing on the input sample set, output the heart failure risk score within the corresponding time window, set multiple risk thresholds according to the heart failure risk score, and when any heart failure risk score exceeds the set risk threshold, trigger an early warning process, output an early warning signal and send an early warning message to the patient terminal device and the medical platform, and at the same time record the early warning event and processing result as feedback information; S5. Compare the feedback information with the heart failure risk score, update the input sample set, and periodically re-execute the feature evaluation and screening processes described in steps S2 and S3 to form a dynamically optimized risk scoring mechanism.

[0008] Optionally, the multi-source physiological and behavioral data specifically includes individual vital sign and behavioral parameter data such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, body temperature, weight, heart rate variability, sleep status, and body movement level collected through Internet of Things devices.

[0009] Optionally, the preprocessing of the multi-source physiological and behavioral data specifically includes performing data cleaning, missing value filling, normalization processing, sliding window segmentation, and feature format unification.

[0010] Optionally, S2 specifically includes: S21. The input sample set is , where represents the original feature vector of the th sample, represents the corresponding heart failure risk label, is the number of samples; S22. Construct a graph structure relationship graph between the input features , where the node set represents the input features, and the edge set represents the correlation edges between the features, and the edge weights are calculated based on the Pearson correlation coefficient between the features; S23. Use the graph convolutional network to perform graph structure embedding processing on each sample in the structured input sample set to generate a graph structure embedding vector that fuses feature dependency relationships ; S24. Concatenate the graph structure embedding vector and the original feature vector to form an enhanced input vector , and construct an enhanced input sample set ; S25. Based on the enhanced input sample set, construct an extreme gradient boosting tree algorithm model. The extreme gradient boosting tree algorithm model consists of multiple regression trees. Use the enhanced input vector as the input of the extreme gradient boosting tree algorithm model, and continuously accumulate the outputs of each regression tree through iterative training to gradually approximate the heart failure risk label value of the corresponding sample; S26. Define an objective function of the third-order approximation based on the enhanced input vector ; S27. Construct the structural regularization term in the objective function ; S28. Set the initial predicted value of all samples to a constant term , and set the initial learning rate to ; S29. In the th round of training, comprehensively consider the current model residual decline rate , the cumulative prediction error fluctuation , and the dynamic interval width of the heart failure risk score result in the previous round , dynamically calculate the adaptive learning rate , and adaptively control the update step size during the training of the extreme gradient boosting tree algorithm model; S210. Based on the objective function and the structural regularization term, construct a regression tree in the current round using a greedy splitting strategy , and update the predicted output of the current extreme gradient boosting tree algorithm model with an adaptive learning rate ; S211. Repeat steps S25 to S210 until the extreme gradient boosting tree algorithm model reaches the set upper limit of the number of regression trees or the objective function satisfies the convergence condition; S212. Calculate the feature importance score of each input feature based on the gain contribution degree of each input feature in the trained extreme gradient boosting tree algorithm model to the objective function , and output a set of feature importance scores , where M is the feature dimension.

[0011] Optionally, the specific steps of S3 are as follows: S31. Receive an input sample set , where represents the original feature vector of the -th sample, represents the corresponding heart failure risk label, is the number of samples; S32. According to the output set of feature importance scores , introduce a feature distribution consistency weight coefficient , and calculate a weighted score vector , where the feature distribution consistency weight coefficient is obtained by calculating the distribution difference degree of the feature in each subsample set and normalizing it: ; where M is the feature dimension; S33. Based on the weighted score vector , perform an ascending order sorting, select the features with the lowest scores, construct a set of indices of features to be removed , and remove the features corresponding to the set of indices of features to be removed from the input sample set to construct a sample set of the first-round feature subset ; S34. Input the sample set of the first-round feature subset into the extreme gradient boosting tree algorithm model for training, and output the corresponding feature importance score vector ; S35. For each feature in the set of indices of features to be removed , calculate the marginal impact on the objective function of the extreme gradient boosting tree algorithm model ; S36. Set the current iteration round as , and based on the feature subset sample set of the previous round and the feature importance scoring vector , perform the following steps: Calculate the feature retention score according to the marginal impact ; Combine the feature retention score and the feature distribution consistency weight coefficient, perform a hierarchical elimination strategy, and construct a new round of feature subset sample set ; Input the new round of feature subset sample set into the extreme gradient boosting tree algorithm model for training, and output the corresponding feature importance scoring vector ; Calculate the performance metrics of the extreme gradient boosting tree algorithm model in this round ; S37. Determine whether the iteration termination condition is satisfied. If it is satisfied, terminate the recursive feature elimination process; otherwise, let , and return to step S36 to continue execution. The iteration termination conditions include: The input dimension in the current feature subset sample set is less than the set minimum feature dimension ; The performance degradation of the extreme gradient boosting tree algorithm model satisfies , where is the performance tolerance threshold; S38. Record the iteration termination round as , extract the set of currently retained feature indices , and use it as the final optimal feature subset; S39. According to the set of feature indices , perform feature mapping on all samples in the structured input sample set to generate the final input sample set .

[0012] Optionally, the specific content of S4 includes: S41. Receive the final input sample set , where represents the input feature vector of the -th sample under the selected optimal feature subset dimension, represents the corresponding heart failure risk label, is the number of samples; S42. For each sample in the final input sample set, the optimal feature vector Input into the constructed and trained extreme gradient boosting tree algorithm model, perform the classification and discrimination process of heart failure risk, and output the heart failure risk score value of each sample within the set time window , which constitutes the heart failure risk score result; S43. Construct a risk score vector from the heart failure risk score values corresponding to all samples , where represents the heart failure risk score result of the th sample in the classification and discrimination process; S44. Set a multi-level risk threshold set , where, satisfying , represents the lower threshold score corresponding to the th risk level, and a total of risk level intervals are constructed; S45. For any score in the risk score vector, determine the risk level interval it belongs to in the set multi-level risk threshold set . There is a unique risk level label that satisfies the inequality , and use the risk level label as the heart failure risk level output of the th sample; S46. Compose the risk level labels corresponding to all samples into a risk level label vector , where represents the heart failure risk level result corresponding to the input feature vector ; S47. Set the risk level threshold for warning trigger . When any risk level label , immediately trigger the warning process and start the heart failure risk emergency response program; S48. Compose the sample indices that meet the warning conditions into a warning trigger sample set , and send a warning signal and a warning information packet containing the risk level label, risk score value, and time window information to the corresponding patient terminal device and medical service platform for each sample in the warning trigger sample set ; S49. Record the sample number , heart failure risk score value , risk level label , the patient terminal device number sent, and the warning response processing information returned by the medical platform for each warning event, and construct feedback information .

[0013] Optionally, S5 specifically includes: S51. Extract the sample index, heart failure risk score result, risk level label, and actual response processing result of historical warning events from the feedback information to construct a feedback sample set; S52. Compare the score results in the feedback sample set with the corresponding feedback information, analyze the scoring accuracy of the extreme gradient boosting tree algorithm model and the rationality of risk level division, and identify samples and feature attributes with a scoring deviation greater than a preset deviation threshold; S53. According to the comparison and analysis results, correct the relevant sample labels in the original input sample set, and supplement new feedback samples and their latest label information to form an updated input sample set; S54. Use the updated input sample set as input data, and re - execute the extreme gradient boosting tree algorithm model construction and training process described in step S2 to obtain a new feature importance score; S55. Based on the new feature importance score, re - execute the recursive feature elimination process described in step S3 to output a new optimal feature subset; S56. Use the updated input sample set and the new optimal feature subset as the basis for the next - cycle heart failure risk scoring process to achieve dynamic optimization and adaptive iteration of the risk scoring mechanism.

[0014] The beneficial effects of the present invention are: By deeply integrating the extreme gradient boosting tree algorithm and the recursive feature elimination method, the present invention constructs an intelligent early - warning method suitable for remote monitoring of heart failure patients, significantly improving the accuracy of heart failure risk assessment and the timeliness of system response. Compared with the problems in the prior art of lacking dynamic optimization of feature selection and the model being easily interfered by redundant features, the present invention introduces an iterative elimination mechanism based on feature importance scoring at the feature level, effectively screening out the most valuable feature subset for risk discrimination, reducing the model complexity and enhancing the generalization ability. At the same time, in the model training process, by introducing a third - order derivative acceleration convergence mechanism, adaptive learning rate regulation, graph - structure feature embedding, and structure regularization optimization, the model stability and training efficiency are further improved.

[0015] The present invention establishes a multi - level risk threshold division and feedback comparison mechanism, which can not only finely warn high - risk events, but also record the actual processing results and use them for model update, constructing a dynamic closed - loop between risk scoring and actual clinical feedback. Through the continuous accumulation and label update of feedback samples, the model can continuously adaptively adjust during operation, and has robustness and evolution ability in the long - term deployment scenario. The overall system integrates functional modules such as data acquisition, feature engineering, risk scoring, early - warning response, and model retraining, forming a complete link covering "perception - judgment - response - optimization", and is suitable for continuous monitoring and active intervention in the remote health management scenario.

[0016] While ensuring the accuracy of heart failure risk identification, the present invention significantly improves the interpretability, self - adaptability and system practicability of the model, and has good medical auxiliary value and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings: Figure 1 is a schematic structural diagram of a remote monitoring and warning system for heart failure patients based on the Internet of Things proposed by the present invention; Figure 2 is a schematic flow diagram of a remote monitoring and warning method for heart failure patients based on the Internet of Things proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0019] Refer to Figure 1 , a remote monitoring and warning system for heart failure patients based on the Internet of Things includes the following modules: A data acquisition and pre - processing module, which is used to acquire multi - source physiological and behavioral data and perform pre - processing to generate an input sample set; A feature evaluation module, which is used to construct an extreme gradient boosting tree algorithm model and train it to obtain feature importance scores; A feature screening module, which is used to execute a recursive feature elimination process and output an optimal feature subset; A risk discrimination module, which is used to input the optimal feature subset into the extreme gradient boosting tree algorithm model to generate a heart failure risk score; An early warning trigger module, which is used to compare the heart failure risk score with a preset risk threshold, trigger an early warning and send an early warning message to the patient and the platform; A feedback recording module, which is used to record the early warning score, early warning level and platform response result to generate feedback information; A dynamic optimization module, which is used to update the input sample set based on the feedback information and retrain the extreme gradient boosting tree algorithm model and screen features regularly.

[0020] Refer to Figure 2 , a remote monitoring and warning method for heart failure patients based on the Internet of Things includes the following steps: S1. Collect multi - source physiological and behavioral data of heart failure patients through Internet of Things devices, perform pre - processing on the multi - source physiological and behavioral data, and construct a structured input sample set; S2. Build an extreme gradient boosting tree algorithm model, train it through the input sample set, and obtain the feature importance scores corresponding to each input feature; S3. Based on the feature importance scores, perform a recursive feature elimination process. In each iteration, eliminate the feature with the lowest score and retrain the extreme gradient boosting tree algorithm model until the set optimal feature subset size or performance convergence condition is met, and output the optimal feature subset; S4. Based on the optimal feature subset, perform discrimination processing on the input sample set, output the heart failure risk score within the corresponding time window, set multi-level risk thresholds according to the heart failure risk score. When any heart failure risk score exceeds the set risk threshold, trigger the warning process, output a warning signal and send a warning message to the patient terminal device and the medical platform, and at the same time record the warning event and the processing result as feedback information; S5. Compare the feedback information with the heart failure risk score, update the input sample set, and regularly re-execute the feature evaluation and screening processes described in steps S2 and S3 to form a dynamically optimized risk scoring mechanism.

[0021] The remote monitoring and warning method for heart failure patients based on the fusion of extreme gradient boosting tree and recursive feature elimination provided by the present invention has significant beneficial effects. This method collects multi-source physiological and behavioral data of patients through Internet of Things devices and constructs a structured input sample set to achieve efficient management and organization of complex high-dimensional health data. Using the extreme gradient boosting tree algorithm to score the importance of each feature and combining the recursive feature elimination process for iterative screening, effectively eliminates redundant and low-contribution features, and improves the discrimination ability and operation efficiency of the model. The constructed heart failure risk scoring mechanism has a hierarchical warning function, can automatically trigger corresponding warning responses according to different risk levels, and improves the timeliness and active intervention ability of remote monitoring. By introducing a comparison mechanism between feedback information and scoring results, the continuous update and optimization of the model input samples can be realized, supporting the long-term adaptation and dynamic learning of patient state changes, and enhancing the robustness and evolution ability of the model. The overall method realizes a full-process closed-loop from data collection, feature optimization, risk discrimination to dynamic feedback, and has the advantages of strong interpretability, high prediction accuracy, fast warning response and system sustainable optimization, and is suitable for remote health monitoring scenarios in chronic disease management.

[0022] In this embodiment, the multi-source physiological and behavioral data specifically includes individual vital sign and behavioral parameter data of heart rate, blood pressure, respiratory rate, blood oxygen saturation, body temperature, weight, heart rate variability, sleep state and body movement level collected through Internet of Things devices.

[0023] In this embodiment, the preprocessing of multi-source physiological and behavioral data specifically includes performing data cleaning, missing value filling, normalization, sliding window segmentation, and feature format unification. The execution of data cleaning, missing value filling, normalization, sliding window segmentation, and feature format unification specifically means performing data cleaning on the multi-source physiological and behavioral data collected by IoT devices, and removing data entries containing outliers, duplicate records, or incorrect formats; then performing missing value filling, and for feature items with missing measurements, reasonably completing them according to the time series context or statistical methods; then performing normalization to convert feature data with different dimensions to a unified numerical range to eliminate the impact of dimensional differences on model training; subsequently performing sliding window segmentation, and dividing the original time series into multiple sample segments according to the set time window length and step size, with each segment being an independent input sample; finally, performing feature format unification, aligning and resampling the data from different devices and different frequencies to unify them into a feature representation format with consistent dimensions and structures, forming a standardized input sample set.

[0024] In this embodiment, S2 specifically includes: S21. The input sample set is , where represents the original feature vector of the th sample, represents the corresponding heart failure risk label, is the number of samples; S22. Construct a graph structure relationship graph between input features , where the node set represents the input features, and the edge set represents the correlation edges between features, and the edge weights are calculated based on the Pearson correlation coefficient between features; S23. Use a graph convolutional network to perform graph structure embedding processing on each sample in the structured input sample set to generate a graph structure embedding vector that fuses feature dependency relationships. The graph structure embedding processing refers to, for the original feature vector of each sample in the structured input sample set, combining the pre-constructed feature graph structure relationship graph, and using the graph convolutional network to perform graph convolution operations on feature nodes, fully fusing the structural dependency relationships between each feature and its adjacent features, and generating a graph structure embedding vector containing global structural information; S24. Concatenate the graph structure embedding vector and the original feature vector to form an enhanced input vector , and construct an enhanced input sample set ; S25. Based on the enhanced input sample set, construct an Extreme Gradient Boosting (XGBoost) tree algorithm model. The XGBoost tree algorithm model consists of multiple regression trees. Using the enhanced input vector as the input of the XGBoost tree algorithm model, through iterative training, continuously accumulate the outputs of each regression tree, and gradually approximate the heart failure risk label value of the corresponding sample. S26. Define the objective function of the third-order approximation based on the enhanced input vector : ; where , , respectively represent the first-order, second-order, and third-order derivatives of the loss function with respect to the model prediction value. represents the structural regularization term in the objective function. is the output value of the t-th regression tree for the enhanced input vector of the i-th sample. represents the t-th regression tree function generated in the t-th round of training. The objective function , its practical significance lies in improving the convergence efficiency and prediction accuracy of the XGBoost tree algorithm in dealing with heart failure risk prediction tasks. This objective function not only introduces first-order and second-order derivative information to capture the direction and curvature of the prediction error, but also further introduces the third-order derivative, thereby enhancing the model's ability to understand the shape of the objective function, optimizing the construction direction and step size of each round of regression tree, and improving the fineness and convergence speed of gradient update. In medical tasks, especially in the highly sensitive human health monitoring scenario, minor changes in prediction errors may lead to serious consequences. Therefore, introducing the third-order derivative to enhance the response ability of the objective function to minor error fluctuations has important practical value. The structural regularization term is incorporated into this objective function to constrain the complexity of the tree structure and prevent overfitting, thus ensuring the generalization performance while enhancing the model's expressive ability. Overall, this third-order objective function significantly improves the fine modeling ability of the model on the basis of maintaining the high efficiency advantage of the traditional XGBoost tree, making it more suitable for clinical decision-making tasks such as heart failure risk scoring.

[0025] S27. Construct the structural regularization term in the objective function : ; where represents the number of leaf nodes of the -th regression tree. is the prediction weight of the -th leaf node. represents the information entropy value calculated based on the number of samples of each leaf node. , , is the structural regularization coefficient; structural regularization term , whose practical significance lies in effectively constraining the complexity of the extreme gradient boosting tree algorithm model, preventing overfitting in the model training process, and thus improving its generalization ability and stability in the heart failure risk prediction task. This regularization term comprehensively considers three aspects of factors: First, by restricting the number of leaf nodes of the tree, the structural scale of the model is controlled to avoid the complex expansion of the model caused by excessive splitting; Second, the square penalty of the prediction weight of the leaf node is introduced to prevent a single node from causing excessive fluctuations in the prediction result; Third, the information entropy term based on the sample distribution is used to measure the partition balance of the leaf nodes, guiding the model to generate a more reasonable structure during the splitting process, thereby enhancing the robustness and interpretability of the model. In the scenario of remote monitoring and early warning for heart failure patients, the model not only requires accurate prediction but also good stability and fault tolerance for boundary samples. Therefore, this structural regularization term plays a crucial role in balancing and regulating in the objective function, ensuring that the model achieves an optimal compromise between performance and complexity and adapting to the dual requirements of reliability and real-time of the actual medical early warning system.

[0026] S28. Set the initial prediction value of all samples to a constant term , and set the initial learning rate to ; S29. In the th round of training, comprehensively consider the current model residual decline rate , the cumulative prediction error fluctuation , and the dynamic interval width of the heart failure risk score result in the previous round , dynamically calculate the adaptive learning rate , and adaptively control the update step size during the training process of the extreme gradient boosting tree algorithm model: ; where is the initial learning rate, , , are regulation factors; adaptive learning rate Based on three core factors, namely, the residual descent rate of the current model, the fluctuation degree of the cumulative prediction error, and the dynamic interval width of the previous heart failure risk score result, the performance change of the model in each round of training is evaluated in real time, and then the update step size is dynamically regulated. Specifically, when the model shows a stable error descent trend and small prediction fluctuations, the learning rate is appropriately increased to accelerate convergence; on the contrary, when the error fluctuations are large or the prediction is unstable, the learning rate is decreased to avoid oscillations and overfitting. This mechanism is significantly different from the traditional fixed learning rate strategy and can flexibly adjust the optimization pace according to the characteristics of different stages in the training process. In the scenario of remote monitoring and early warning of heart failure patients based on the Internet of Things, the data distribution has individual differences and time variability. This adaptive strategy ensures that the model can continuously adapt to the learning needs under different data scenarios, thereby enhancing the accuracy and robustness of the risk score result and improving the intelligent response ability of the overall system.

[0027] S210. Based on the objective function and the structural regularization term, construct a regression tree using the greedy splitting strategy in the current round , and use the adaptive learning rate to update the prediction output of the current extreme gradient boosting tree algorithm model; S211. Repeat steps S25 to S210 until the extreme gradient boosting tree algorithm model reaches the set upper limit of the number of regression trees or the objective function satisfies the convergence condition; S212. Calculate the feature importance score of each input feature based on the gain contribution degree of each input feature in the trained extreme gradient boosting tree algorithm model to the objective function , and output the set of feature importance scores , where M is the feature dimension.

[0028] The present invention structurally improves the extreme gradient boosting tree algorithm model by introducing a graph structure embedding mechanism and a multi-order derivative optimization strategy, significantly enhancing the expression ability and learning efficiency of the model in the heart failure risk prediction task. The graph convolutional network is used to extract the structural correlation information between input features, forming a graph structure embedding vector that fuses feature relationships, effectively overcoming the limitations of traditional feature independent assumptions and improving the modeling ability for complex feature interaction relationships. By concatenating the graph structure embedding vector and the original features, an enhanced input sample set is constructed, improving the information density and discriminative ability of the model input. The introduction of the third-order derivative to construct an approximate objective function can more accurately capture the gradient change trend and accelerate the model convergence speed; the constructed structural regularization term effectively suppresses the overfitting risk and enhances the model generalization performance. During the model training process, the learning rate is dynamically adjusted through the joint judgment of the residual descent rate, the prediction error fluctuation degree, and the dynamic range of the risk score, realizing the adaptive control of the model training step size. The feature importance score is output based on the gain contribution degree of the feature to the objective function, providing efficient support for subsequent feature screening and risk discrimination. The overall method combines the modeling advantages of graph neural networks with the strong generalization ability of gradient boosting trees, having the beneficial effects of accurate modeling, efficient training, and controllable structure.

[0029] In this embodiment, the S3 specifically includes: S31. Receive the input sample set , where represents the original feature vector of the -th sample, represents the corresponding heart failure risk label, is the number of samples; S32. According to the output feature importance score set , introduce the feature distribution consistency weight coefficient , and calculate the weighted score vector , where the feature distribution consistency weight coefficient is obtained by calculating the distribution difference degree of the feature in each sub-sample set and normalizing it: ; where, M is the feature dimension; S33. Based on the weighted score vector , perform ascending sorting, select the features with the lowest scores, construct the set of indices of features to be excluded, and exclude the features corresponding to the set of indices of features to be excluded from the input sample set to construct the first-round feature subset sample set ; S34. Use the first-round feature subset sample set Input into the extreme gradient boosting tree algorithm model for training, and output the corresponding feature importance score vector ; S35. Feature index set to be removed Each feature in , calculate the marginal impact on the objective function of the extreme gradient boosting tree algorithm model , the calculation of the marginal impact on the objective function of the extreme gradient boosting tree algorithm model means that in the current round of feature elimination iteration, for each feature in the feature index set to be eliminated, the input sample sets before and after the feature is eliminated are respectively constructed, and they are respectively input into the extreme gradient boosting tree algorithm model for training, and the minimum value of the corresponding objective function is recorded. By comparing the difference in the objective function value before and after the feature is eliminated, the marginal gain or loss of the feature on the overall loss of the model is calculated, thereby measuring the actual impact of the feature on the model performance; S36. Set the current iteration round to , based on the previous round of feature subset sample set and the feature importance score vector , perform the following steps: Calculate the feature retention score based on the marginal influence : ; in, is the feature fusion weight, For the The feature importance score calculated in the extreme gradient boosted tree algorithm model for the jth feature in the round iteration; Feature retention score The practical significance of is to quantify the retention value of each feature to the model performance in the current iteration round, so as to provide a scientific and controllable elimination basis for the recursive feature elimination process. By integrating two key factors - the feature importance score of the feature in the extreme gradient boosting tree algorithm model in the previous round of training and its marginal influence on the model objective function, a "feature retention score" is comprehensively constructed. Among them, the feature importance score reflects the direct contribution of the feature to the model prediction performance, while the marginal influence measures the sensitivity of the objective function to the change after the feature is eliminated. By introducing the fusion weight λ, the relative weights of the two in feature evaluation can be flexibly controlled to adapt to the model optimization needs in different scenarios. This scoring mechanism helps to avoid the loss of important information caused by over-reliance on a single evaluation indicator, and more accurately distinguishes "redundant features" from "key features" in the feature elimination stage. Especially in the remote monitoring and early warning system for heart failure patients, the features come from a wide range of sources and are highly heterogeneous, providing a multi-dimensional and explainable feature retention evaluation method, providing a reliable basis for building a streamlined and efficient model, and effectively improving the stability and accuracy of the final risk scoring results.

[0030] Execute a hierarchical elimination strategy by combining the feature retention score and the feature distribution consistency weight coefficient to construct a new round of feature subset sample sets ; Input the new round of feature subset sample sets into the extreme gradient boosting tree algorithm model for training, and output the corresponding feature importance score vector ; Calculate the performance metrics of the extreme gradient boosting tree algorithm model in this round , and calculating the performance metrics of the extreme gradient boosting tree algorithm model in this round is based on the prediction performance of the model on the validation set or cross-validation subset after the current round of training. Common evaluation metrics include mean absolute error, mean squared error, precision, recall, or F1 value. The specific approach is as follows: Apply the trained model to the validation data to obtain the error between the model prediction output and the true label, and calculate the performance value of the current round through the selected metric formula; at the same time, the performance metrics of the previous round can be compared to determine whether the model performance has improved, whether the convergence standard is reached, or the performance degradation tolerance threshold is satisfied to assist in whether the iterative process continues.

[0031] S37. Determine whether the iteration termination condition is satisfied. If it is satisfied, terminate the recursive feature elimination process; otherwise, let , and return to step S36 to continue execution. The iteration termination conditions include: The input dimension in the current feature subset sample set is less than the set minimum feature dimension ; The performance degradation of the extreme gradient boosting tree algorithm model satisfies , where is the performance tolerance threshold; S38. Record the iteration termination round as , extract the currently retained feature index set , and use it as the final optimal feature subset; S39. According to the feature index set , perform feature mapping on all samples in the structured input sample set to generate the final input sample set .

[0032] The present invention introduces a feature distribution consistency evaluation and marginal impact analysis mechanism into the recursive feature elimination process, significantly improving the scientific nature of feature screening and the stability of model performance. By constructing a feature distribution consistency weight coefficient to measure the distribution stability of features in different subsample sets, it effectively avoids the fluctuation of importance scores caused by sample perturbation and enhances the robustness of feature screening. The introduction of a weighted scoring vector combines the dual dimensions of feature importance and distribution consistency to ensure that the prediction contribution and distribution stability of features are considered during the process of feature elimination. Further, by calculating the marginal impact of candidate features to be eliminated on the objective function, the actual impact on model performance after their elimination is evaluated, improving the accuracy of feature elimination decisions. During the iteration process, a hierarchical elimination strategy is executed based on the feature retention score and the distribution consistency weight to gradually construct a high-quality feature subset, and the iteration termination condition is dynamically judged by integrating the model performance evaluation results to ensure that the performance is not lost during the feature compression process. The finally output optimal feature subset effectively compresses the input dimension while maintaining the generalization ability of the model, reduces the system operation overhead, and improves the real-time performance and deployment efficiency of the model. The overall process integrates stability evaluation and performance feedback adjustment, has higher interpretability and adaptability, and is particularly suitable for risk screening scenarios under high-dimensional health monitoring data.

[0033] In this embodiment, step S4 specifically includes: S41. Receive the final input sample set , where represents the input feature vector of the -th sample under the dimension of the selected optimal feature subset, represents the corresponding heart failure risk label, is the number of samples; S42. Input the optimal feature vector of each sample in the final input sample set into the constructed and trained extreme gradient boosting tree algorithm model, execute the classification and discrimination process of heart failure risk, and output the heart failure risk score value of each sample within the set time window, forming a heart failure risk score result. The execution of the classification and discrimination process of heart failure risk specifically means inputting the optimal feature subset vector of each sample in the final input sample set into the trained extreme gradient boosting tree algorithm model, jointly predicting the sample through multiple decision regression trees constructed inside the model, accumulating the output results of each regression tree, generating the heart failure risk score value of the sample within the current time window, and using the heart failure risk score value as the quantitative discrimination result of the heart failure risk degree; S43. Form a risk score vector with the heart failure risk score values corresponding to all samples, where represents the The heart failure risk score results of a sample during the classification and discrimination process; S44. Set a multi-level risk threshold set , where , represents the lower threshold of the score corresponding to the th risk level. A total of risk level intervals are constructed; S45. For any score in the risk score vector, determine the risk level interval it belongs to in the set of multi-level risk thresholds set. There is a unique risk level label that satisfies the inequality . And use the risk level label as the heart failure risk level output of the th sample; S46. Combine the risk level labels corresponding to all samples into a risk level label vector , where represents the heart failure risk level result corresponding to the input feature vector ; S47. Set the risk level threshold for warning trigger. When any risk level label , immediately trigger the warning process and start the heart failure risk emergency response program; S48. Combine the sample indices that meet the warning conditions into a warning trigger sample set . And send a warning signal and a warning information packet containing the risk level label, risk score value, and time window information to the corresponding patient terminal device and medical service platform for each sample in the warning trigger sample set ; S49. Record the sample number , heart failure risk score value , risk level label , the patient terminal device number sent, and the warning response processing information returned by the medical platform for each warning event, and construct feedback information .

[0034] The present invention realizes the refined discrimination and hierarchical evaluation of heart failure risk by constructing an extreme gradient boosting tree classification structure based on the optimal feature subset, significantly enhancing the accuracy and practicability of the early warning system. Using high-quality feature input to improve the discriminative ability of the model for heart failure risk scoring effectively improves the accuracy and stability of the heart failure risk prediction results; setting multi-level risk thresholds and outputting risk level labels realizes the interval division and labeling of the heart failure risk degree, facilitating medical staff to quickly grasp the risk status of patients. Setting an early warning trigger mechanism and response logic, when the risk score exceeds the set threshold, it can automatically identify high-risk individuals and initiate the early warning program, realizing efficient risk response and multi-terminal linkage early warning information push, and improving the emergency response speed and coverage breadth of the system for sudden heart failure events. A complete early warning information recording mechanism is established to ensure that the key data, response actions, and processing results of each early warning event are retained by the system, providing detailed data support for evaluation, optimization, and clinical decision-making. The overall technical solution constructs a closed-loop risk discrimination - hierarchical evaluation - early warning trigger - information push - response feedback chain, while ensuring the prediction performance, strengthening the real-time monitoring and auxiliary diagnosis and treatment capabilities of the system, and is applicable to the intelligent remote monitoring and risk intervention scenarios of large-scale heart failure patients based on the Internet of Things environment.

[0035] In this embodiment, step S5 specifically includes: S51. Extract the sample index, heart failure risk score result, risk level label, and actual response processing result of historical early warning events from the feedback information to construct a feedback sample set; S52. Compare the score results in the feedback sample set with the corresponding feedback information, analyze the scoring accuracy of the extreme gradient boosting tree algorithm model and the rationality of risk level division, and identify the samples and feature attributes with a scoring deviation greater than the preset deviation threshold. The comparison of the score results in the feedback sample set with the corresponding feedback information and the analysis of the scoring accuracy of the extreme gradient boosting tree algorithm model and the rationality of risk level division specifically refer to comparing the heart failure risk score value generated by each historical early warning sample in the model and its corresponding risk level label with the actual clinical response result one by one, judging whether the model score accurately reflects the severity of the condition, and whether the risk level division reasonably triggers an effective early warning, so as to identify the samples with obvious deviations and analyze the corresponding feature attributes and label setting problems; S53. According to the comparison and analysis results, correct the relevant sample labels in the original input sample set, and supplement the newly added feedback samples and their latest label information to form an updated input sample set; S54. Use the updated input sample set as input data, and re-execute the extreme gradient boosting tree algorithm model construction and training process described in step S2 to obtain a new feature importance score; S55. Re - execute the recursive feature elimination process described in step S3 based on the new feature importance score, and output a new optimal feature subset. S56. Use the updated input sample set and the new optimal feature subset as the basis for the next - cycle heart failure risk score processing to achieve the dynamic optimization and adaptive iteration of the risk scoring mechanism.

[0036] The present invention realizes the adaptive update and dynamic optimization of the scoring mechanism by introducing feedback information to construct a feedback sample set. By extracting and analyzing the risk scores and response processing results of historical warning events, samples and key features with large scoring errors can be accurately identified, so as to timely correct the input sample set and label data, supplement new feedback samples, and enhance the timeliness and representativeness of the sample set. Retraining the extreme gradient boosting tree algorithm model and updating the feature importance score helps to construct an optimal feature subset that better fits the current clinical reality. It not only enhances the model's adaptability to new feature patterns but also effectively improves the accuracy of risk scores and the credibility of warning judgments. By constructing a closed - loop optimization mechanism, the heart failure risk identification system can dynamically respond to real - time feedback, realize model self - correction and performance iterative update, and significantly improve the intelligent level of the system and the effect of remote health management.

[0037] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the cardiovascular chronic disease management center of a certain tertiary - level hospital. To improve the quality of discharge follow - up for patients with chronic heart failure, the research team selects 10 typical heart failure patients who were treated and discharged during the period from May 10th to May 24th, 2025, and conducts a remote monitoring pilot study based on the present invention. The average age of the patients is 64 years old, and they all suffer from left ventricular dysfunction above grade 2, with typical symptoms such as paroxysmal nocturnal dyspnea and decreased exercise tolerance.

[0038] All patients are equipped with wearable devices with Internet of Things communication modules before discharge, including continuous electrocardiogram acquisition chest patches, wireless blood pressure monitors, sleep and activity wristbands, and Bluetooth blood oxygen meters. The devices are interconnected with the hospital information platform through home Wi - Fi and regularly upload seven types of multi - source physiological and behavioral data, including heart rate, blood pressure, respiratory rate, blood oxygen saturation, frequency of nocturnal awakening, and exercise volume, every day.

[0039] The platform backend first calls the data preprocessing module to perform standardization, noise removal, and time window alignment, constructing a structured input sample set. Subsequently, this sample set is input into the improved extreme gradient boosting tree model, which integrates graph structure feature embedding and third-order derivative acceleration of the objective function, enhancing the ability to capture non-linear risk factors. After the model outputs the importance scores of each feature, the recursive feature elimination module starts the hierarchical elimination and marginal impact calculation process, generating a 10-dimensional optimal feature subset while ensuring the stable performance of the model.

[0040] On this basis, the model performs risk scoring on the samples uploaded every day, setting the risk threshold at four levels: normal (<0.5), attention (0.5–0.7), high risk (0.7–0.85), and extremely high risk (>0.85). Once the score exceeds 0.85, the platform immediately sends warning messages to the patient APP and the medical staff workstation, and records the event number, warning level, timestamp, and subsequent processing results.

[0041] Table 1 Comparison effect table between the traditional method and the method of the present invention

[0042] From the data in Table 1, it can be clearly observed that the method of the present invention has more excellent effects in the remote monitoring and warning of heart failure patients compared with the traditional method. In the traditional method, only 4 patients (P003, P006, P008, P010) were successfully warned, and their corresponding risk scores were relatively high. However, for the other 6 patients, even though there were certain risk manifestations, they were not identified in time, which was likely to cause potential missed reports and affect the timeliness of timely intervention.

[0043] In contrast, the method of the present invention combines extreme gradient boosting tree and recursive feature elimination technology. After fully exploring the complex correlation relationships between input features, it successfully realizes the accurate identification of all high-risk patients. A total of 8 patients were triggered for warning (including patients such as P001, P002, P005, P007, P009 who were not warned by the traditional method), and all warning results were consistent with the actual clinical intervention results. Taking P001 as an example, the traditional method scored 0.58 and did not trigger a warning, while the method of the present invention identified it as 0.76 and initiated remote diagnosis, successfully intervening in possible emergencies. Another example is P007 and P009. The traditional scores were both lower than 0.5 and did not trigger a warning, but they were respectively identified as 0.65 and 0.63 by the method of the present invention. Finally, the doctor called for follow-up and intervention through the remote system.

[0044] It is worth noting that there is no false alarm in the method of the present invention. For example, patient P004 did not trigger an alarm under both methods, and there was no feedback on risk changes clinically, further indicating that the system has strong stability and robustness. In addition, patients such as P006, P008, and P010 already had high scores themselves, but the score of the present invention was further improved, fully amplifying the danger signal and ensuring the foresight and urgency of medical response.

[0045] In summary, the results fully verify that the method of the present invention is not only significantly superior to traditional means in terms of the sensitivity of risk scoring, but also has higher clinical intervention consistency, which has important practical significance for the active prevention and control, personalized management of heart failure patients and the reasonable scheduling of medical resources in the Internet of Things environment.

[0046] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A remote monitoring and warning system for heart failure patients based on the Internet of Things, characterized in that, It includes the following modules: The data acquisition and preprocessing module is used to acquire multi-source physiological and behavioral data and perform preprocessing to generate an input sample set; The feature evaluation module is used to construct and train an extreme gradient boosting tree algorithm model to obtain feature importance scores; The feature screening module is used to execute a recursive feature elimination process and output an optimal feature subset; The risk discrimination module is used to input the optimal feature subset into the extreme gradient boosting tree algorithm model to generate a heart failure risk score; The early warning trigger module is used to compare the heart failure risk score with a preset risk threshold, trigger an early warning and send an early warning message to the patient and the platform; The feedback recording module is used to record the early warning score, early warning level and platform response result to generate feedback information; The dynamic optimization module is used to update the input sample set based on the feedback information, and regularly retrain the extreme gradient boosting tree algorithm model and screen features.

2. The remote monitoring and warning method for heart failure patients based on the Internet of Things is applied to the remote monitoring and warning system for heart failure patients based on the Internet of Things described in claim 1, and is characterized in that, It includes the following steps: S1. Collect multi-source physiological and behavioral data of heart failure patients through Internet of Things devices, preprocess the multi-source physiological and behavioral data, and construct a structured input sample set; S2. Construct an extreme gradient boosting tree algorithm model, train it through the input sample set, and obtain the feature importance scores corresponding to each input feature; S3. Based on the feature importance scores, execute a recursive feature elimination process, eliminate the feature with the lowest score in each iteration and retrain the extreme gradient boosting tree algorithm model until the set optimal feature subset size or performance convergence condition is met, and output the optimal feature subset; S4. Based on the optimal feature subset, perform discrimination processing on the input sample set, output the heart failure risk score within the corresponding time window, set multiple risk thresholds according to the heart failure risk score, and when any heart failure risk score exceeds the set risk threshold, trigger an early warning process, output an early warning signal and send an early warning message to the patient terminal device and the medical platform, and at the same time record the early warning event and processing result as feedback information; S5. Compare the feedback information with the heart failure risk score, update the input sample set, and regularly re-execute the feature evaluation and screening processes described in steps S2 and S3 to form a dynamically optimized risk scoring mechanism.

3. The remote monitoring and warning method for heart failure patients based on the Internet of Things according to claim 2, wherein The multi-source physiological and behavioral data specifically includes individual vital sign and behavioral parameter data of heart rate, blood pressure, respiratory rate, blood oxygen saturation, body temperature, weight, heart rate variability, sleep status and body movement level collected through Internet of Things devices.

4. The remote monitoring and warning method for heart failure patients based on the Internet of Things according to claim 2, characterized in that, The preprocessing of the multi-source physiological and behavioral data specifically includes performing data cleaning, missing value filling, normalization processing, sliding window segmentation and feature format unification.

5. The remote monitoring and early warning method for heart failure patients based on the Internet of Things according to claim 2, characterized in that, The specific content of S2 includes: S21. The input sample set is , where represents the original feature vector of the -th sample, represents the corresponding heart failure risk label, is the number of samples; S22. Construct a graph structure relationship graph among input features , where the node set represents the input features, and the edge set represents the correlation edges among the features, and the edge weights are obtained by calculating the Pearson correlation coefficients between the features; S23. Use the graph convolutional network to perform graph structure embedding processing on each sample in the structured input sample set, and generate a graph structure embedding vector that fuses feature dependencies ; S24. Concatenate the graph structure embedding vector and the original feature vector to form an enhanced input vector , and construct an enhanced input sample set ; S25. Based on the enhanced input sample set, construct an extreme gradient boosting tree algorithm model. The extreme gradient boosting tree algorithm model consists of multiple regression trees. Using the enhanced input vector as the input of the extreme gradient boosting tree algorithm model, the outputs of each regression tree are continuously accumulated through iterative training to gradually approximate the heart failure risk label value of the corresponding sample; S26. Define a third-order approximate objective function based on the enhanced input vector ; S27. Construct the structural regularization term in the objective function ; S28. Set the initial prediction values of all samples to the constant term , and set the initial learning rate to ; S29. In the round of training, comprehensively consider the current model residual descent rate , the cumulative prediction error fluctuation , and the dynamic interval width of the heart failure risk score result in the previous round , dynamically calculate the adaptive learning rate , and adaptively control the update step size during the training process of the extreme gradient boosting tree algorithm model; S210. Based on the objective function and the structural regularization term, construct a regression tree using a greedy splitting strategy in the current round , and use an adaptive learning rate to update the prediction output of the current extreme gradient boosting tree algorithm model; S211. Repeat steps S25 to S210 until the extreme gradient boosting tree algorithm model reaches the set upper limit of the number of regression trees or the objective function meets the convergence condition; S212. Calculate the feature importance score for each input feature based on the gain contribution degree of each input feature to the objective function in the trained extreme gradient boosting tree algorithm model , and output the set of feature importance scores , where M is the feature dimension.

6. The remote monitoring and warning method for heart failure patients based on the Internet of Things according to claim 5, characterized in that The specific content of S3 includes: S31. Receive an input sample set , where represents the original feature vector of the th sample, represents the corresponding heart failure risk label, is the number of samples; S32. According to the output feature importance score set , introduce the feature distribution consistency weight coefficient , calculate the weighted score vector , where the feature distribution consistency weight coefficient is obtained by calculating the distribution difference degree of the feature appearing in each sub-sample set and normalizing it: ； Among them, M is the feature dimension; S33. Based on the weighted scoring vector Perform ascending sorting, and select the features with the lowest scores to construct a set of indexes of features to be excluded , and exclude the features corresponding to the set of indexes of features to be excluded from the input sample set to construct a sample set of the first-round feature subset ; S34. Input the sample set of the first-round feature subset into the extreme gradient boosting tree algorithm model for training, and output the corresponding feature importance scoring vector ; S35. Treat each feature in the set of feature indices to be excluded and calculate the marginal impact on the objective function of the extreme gradient boosting tree algorithm model ; ​ S36. Set the current iteration round as , and based on the sample set of the feature subset in the previous round and the feature importance scoring vector , perform the following steps: Calculate the feature retention score according to the marginal impact amount ; Execute a hierarchical elimination strategy by combining the feature retention score and the feature distribution consistency weight coefficient to construct a new round of feature subset sample sets ; Input the new round of feature subset sample set into the extreme gradient boosting tree algorithm model for training, and output the corresponding feature importance scoring vector ; Calculate the performance metrics of the extreme gradient boosting tree algorithm model for this round ; S37. Determine whether the iterative termination condition is satisfied. If it is satisfied, terminate the recursive feature elimination process; otherwise, let , return to step S36 to continue execution. The iterative termination conditions include: The input dimension in the current feature subset sample set is less than the set minimum feature dimension ; The degradation amplitude of the performance of the extreme gradient boosting tree algorithm model satisfies , where is the performance tolerance threshold; S38. Denote the iteration termination round as , and extract the currently retained feature index set as the final optimal feature subset; S39. According to the feature index set , perform feature mapping on all samples in the structured input sample set to generate the final input sample set .

7. The remote monitoring and warning method for heart failure patients based on the Internet of Things according to claim 6, characterized in that, Specifically, S4 includes: S41. Receive the final input sample set , where represents the input feature vector of the th sample under the dimension of the selected optimal feature subset, represents the corresponding heart failure risk label, is the number of samples; S42. Input the optimal feature vector of each sample in the final input sample set into the constructed and trained extreme gradient boosting tree algorithm model, execute the classification and discrimination process of heart failure risk, and output the heart failure risk score value of each sample within the set time window , forming the heart failure risk score result; S43. Construct a risk score vector from the heart failure risk score values corresponding to all samples , where represents the heart failure risk score result of the th sample in the classification and discrimination process; S44. Set a multi-level risk threshold set , where it satisfies , represents the lower threshold score corresponding to the th risk level, and a total of risk level intervals are constructed; S45. For any score in the risk score vector , determine the risk level interval to which it belongs in the set of multi-level risk thresholds . There is a unique risk level label that satisfies the inequality . Use the risk level label as the output of the heart failure risk level for the th sample; S46. Compose a risk level label vector from the risk level labels corresponding to all samples , where represents the heart failure risk level result corresponding to the input feature vector ; S47. Set the risk level threshold for warning trigger When any risk level label is present, the warning process is immediately triggered and the heart failure risk emergency response program is initiated; S48. Compose a warning-triggering sample set by using the sample indexes that meet the warning conditions , and for each sample's corresponding patient terminal device and medical service platform in the warning-triggering sample set , send a warning signal and a warning information packet that contain a risk level label, a risk score value, and time window information; S49. Record the sample number corresponding to each warning event , heart failure risk score value , risk level label , the patient terminal device number sent and the warning response processing information returned by the medical platform, and construct feedback information .

8. The remote monitoring and warning method for heart failure patients based on the Internet of Things according to claim 7, characterized in that, Specifically, S5 includes: S51. Extract the sample index, heart failure risk score result, risk level label, and actual response processing result of historical warning events from the feedback information to construct a feedback sample set; S52. Compare the score results in the feedback sample set with the corresponding feedback information, analyze the scoring accuracy of the extreme gradient boosting tree algorithm model and the rationality of risk level division, and identify samples and feature attributes with a scoring deviation greater than the preset deviation threshold; S53. According to the comparison and analysis results, correct the relevant sample labels in the original input sample set, and supplement new feedback samples and their latest label information to form an updated input sample set; S54. Use the updated input sample set as input data, and re-execute the extreme gradient boosting tree algorithm model construction and training process described in step S2 to obtain a new feature importance score; S55. Based on the new feature importance score, re-execute the recursive feature elimination process described in step S3 to output a new optimal feature subset; S56. Use the updated input sample set and the new optimal feature subset as the basis for the next cycle of heart failure risk score processing to achieve dynamic optimization and adaptive iteration of the risk scoring mechanism.

Images