Method and system for constructing coronary intervention postoperative risk prediction model

Abstract

The invention discloses a coronary intervention postoperative risk prediction model construction method and system, and belongs to the technical field of medical care information and health monitoring. A coronary intervention postoperative risk prediction model construction method comprises the steps of constructing an attending doctor experience scoring system and a nursing personnel ability scoring system, and performing weighted fusion on results of a doctor experience model and a nursing personnel model to form a final additional model for output. According to the coronary intervention postoperative risk prediction model construction method and system provided by the invention, the main model is combined with the additional model, the main model is comprehensively modeled through the regression model and the time sequence model based on the static and dynamic characteristic data of the patient, and the individual health state change of the patient is captured; the additional model evaluates the influence of medical teams and relatives on the postoperative risk by quantifying the experience of doctors and the ability of caregivers, and overcomes the defect that a traditional model only depends on patient data.

Description

Technical Field

[0001] The present invention relates to the technical field of healthcare information and health monitoring, and specifically to a method and system for constructing a risk prediction model after coronary intervention. Background Art

[0002] Coronary intervention (such as percutaneous coronary intervention, PCI) is an important means for treating coronary heart disease. However, patients may still face the risk of adverse events such as restenosis, myocardial infarction, heart failure, etc. after the operation. Traditional risk assessment methods mostly rely on doctors' experience and simple clinical scoring systems, and their prediction accuracy is limited, making it difficult to meet the needs of personalized medicine. With the development of medical big data and artificial intelligence technologies, it has become possible to construct a precise risk prediction model using machine learning.

[0003] Chinese Patent with publication number CN113314225A discloses a method and device for establishing a postoperative risk prediction model. Based on patient data, a postoperative risk assessment prediction model is constructed through multiple regression analysis to evaluate the risk of adverse events occurring after Stanford type A aortic dissection surgery, improving the survival rate of patients and reducing the complication rate. This application uses the patient's hospital physical examination information as all the basic data for model construction, which can be said to be all existing data, and all existing data are based on the patient as the entire risk subject. However, in real life, patients need to be cared for by family members or medical staff, and the medical staff and caregivers around them also have a greater impact on the postoperative risk of patients. Therefore, there is an urgent need for an efficient and accurate method and system for constructing a risk prediction model after coronary intervention. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for constructing a risk prediction model after coronary intervention. By combining a main model and an additional model, the main model captures the individualized health status changes of patients based on static and dynamic characteristic data of patients, and the additional model evaluates the impact of the medical team and relative care on the postoperative risk by quantifying doctors' experience and caregivers' capabilities, making up for the deficiency of traditional models that only rely on patient data, and solving the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A method for constructing a risk prediction model after coronary intervention, comprising the following steps:

[0006] S1. Build the main model: Obtain the clinical data and life characteristic data of the patient, process the data, divide it into static feature data and dynamic feature data. Based on the static feature data, complete the construction of the regression model through machine learning algorithms, divide the training set and the test set, use cross-validation to optimize the model parameters, construct a time series model using the dynamic feature data, splice the features extracted from the time series model with the static feature data into a new feature vector, retrain the regression model using the comprehensive feature vector, evaluate the model performance, and evaluate the model performance using indicators such as accuracy, recall rate, F1 score, and the area under the ROC curve AUC;

[0007] S2. Build the additional model: Construct the attending physician experience scoring system and the caregiver ability scoring system, predict the postoperative risk of the patient according to the doctor's experience and the caregiver's ability score, and weighted fuse the results of the doctor experience model and the caregiver model to form the final additional model output;

[0008] S3. Integrate the main model and the additional model: Non-linearly fuse the output results of the main model and the additional model to generate the final postoperative risk prediction value.

[0009] Preferably, the criteria for dividing the static feature data and the dynamic feature data are the time scale and the acquisition frequency. The static feature data represents data that remains stable over a period of time and does not change significantly in the short term, usually reflecting the basic attributes of the patient or the long-term health status, including but not limited to the patient's personal information, such as age, gender; physiological indicators that are not easily changed, such as past medical history, genetic factors; long-term habits in lifestyle, such as eating habits, smoking history. The dynamic feature data represents data that fluctuates over time, usually reflecting the health status or behavior pattern of the patient within a specific time period, for example: real-time monitoring data, such as heart rate, blood pressure, blood sugar level; short-term behavior or environmental changes, such as recent exercise frequency, psychological stress level.

[0010] Preferably, the specific steps of S1 to build the main model are as follows:

[0011] S11: Collect the static feature data of the patient, process the missing values, detect and process the outliers, encode the non-numerical data, and normalize the numerical data to ensure that different features have the same dimension;

[0012] S12: Calculate the correlation coefficient between each feature and the postoperative risk, preliminarily screen the features with higher correlation, use the LASSO regression algorithm to automatically screen important features through L1 regularization, and use the random forest algorithm to select key features according to the importance score of the features;

[0013] S13: Divide the training set according to the ratio of 70%-80%, and use the remaining part as the test set. Use stratified sampling to ensure the consistent distribution of the training set and the test set;

[0014] S14: Select a machine learning algorithm to construct a regression model. Use the training set to train the selected algorithm, use K-fold cross-validation to optimize the model parameters, adjust the hyperparameters to improve the model performance, evaluate the performance of the regression model. The evaluation metrics include accuracy, recall, F1-score, and the area under the ROC curve (AUC). Compare the performances of different algorithms and select the optimal model;

[0015] S15: Process the dynamic feature data. Select a time series model, conduct model training, input the time series data, output the predicted values of the time series model, extract the key features of the time series model, evaluate the performance of the time series model, compare the performances of different models, and select the optimal model;

[0016] S16: Concatenate the dynamic features extracted by the time series model with the static feature data, and use the regression model to model this feature vector to generate the final postoperative risk prediction value;

[0017] Preferably, the processing method of the dynamic feature data is as follows: convert the format of the dynamic feature data into a two-dimensional array, where each row represents a time point and each column represents a variable. For missing values, interpolation can be used to fill them. Use a rolling window statistic to detect outliers, correct or delete the outliers, and normalize or standardize the numerical data to ensure that different features have the same dimension.

[0018] Preferably, the process of setting the corresponding duration of the rolling window includes:

[0019] Extract the time scale corresponding to each dynamic feature data;

[0020] Use the time scale corresponding to each dynamic feature data to obtain the time scale difference between every two dynamic feature data;

[0021] Compare the time scale difference between the two dynamic feature data with a preset time scale difference threshold;

[0022] When the time scale difference between every two dynamic feature data included in all dynamic feature data is less than the preset time scale difference threshold, then use the maximum value of the time scale of the corresponding dynamic feature data in all dynamic feature data as the corresponding duration of the rolling window;

[0023] When the time scale difference between every two dynamic feature data contained in all dynamic feature data is not less than the preset time scale difference threshold, the corresponding time length of the rolling window is set by using the time scale corresponding to each dynamic feature data and the number of data collection times of each dynamic feature data within the time scale.

[0024] Preferably, the time scale corresponding to each dynamic feature data is combined with the number of data collection times of each dynamic feature data within the time scale to set the corresponding duration of the rolling window, including:

[0025] The time scale for extracting each dynamic feature data;

[0026] Extract the data collection times of each dynamic feature data within the time scale;

[0027] The window setting coefficient corresponding to each dynamic feature data is obtained by using the time scale corresponding to each dynamic feature data and the number of data collection times of each dynamic feature data within the time scale;

[0028] The window setting coefficient corresponding to each dynamic feature data is obtained by the following formula:

[0029]

[0030] Wherein, W(T, N) represents the window setting coefficient corresponding to each dynamic feature data. The larger the W(T, N) value is, the more drastic the dynamic change of the data is (or the more complex the characteristics are), and a more flexible window setting is required (such as shortening the window length or increasing the update frequency); the smaller the W(T, N) value is, the weaker the data dynamics is or the sampling is sparse, and the current window length should be appropriately increased; N represents the number of data collection times for each dynamic feature data within the time scale; T represents the time scale of each dynamic feature data; α represents the dynamic sensitivity factor, which is used to adjust the sensitivity of the collection corresponding to the change of dynamic feature data. The value range is 0.7-1.2, preferably 1.0; β represents the preset baseline sampling rate, which is used to represent the standard number of samplings per unit time; r represents the time decay constant, with a value range of 0.3-0.7, preferably 0.5, which is used to control the attenuation effect of the time scale on the window; T0 represents the preset reference time scale, which is used to normalize the time dimension; k represents the nonlinear correction index, which is used to adjust the nonlinear influence of the time scale, with a value range of 1.5-2.3, preferably 2; the above technical solution automatically adapts to the second-level to hour-level data through the modulation term to avoid manual segmentation settings. The logarithmic term suppresses high-frequency sampling interference, and the exponential term filters invalid time scales. The parameter open interface supports integration with the AIoT platform to meet the dynamic optimization needs of Industry 4.0. This formula significantly surpasses the traditional linear model in terms of functional coupling and parameter interpretability, and meets the requirements of patent law for creativity.

[0031] By Quantify the data acquisition density, compress the marginal benefit of high sampling rate through natural logarithm, introduce logarithmic function to avoid numerical explosion caused by linear growth, and retain the sensitivity of density difference. Suppress the window setting weights for ultra-long or ultra-short time scales, prevent extreme values from interfering, combine exponential decay function with power-law correction to strengthen the penalty for unconventional time scales;

[0032] Compare the window setting coefficient corresponding to each of the dynamic feature data with a preset window setting coefficient threshold;

[0033] When the window setting coefficient corresponding to each of the dynamic feature data is lower than the preset window setting coefficient, then the time scales of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times are combined with the window setting coefficient corresponding to the dynamic feature data to obtain the corresponding duration of the rolling window;

[0034] Among them, the corresponding duration of the rolling window is obtained through the following formula:

[0035]

[0036] Among them, T k represents the corresponding duration of the rolling window; T max and T min respectively represent the time scales of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times; W max and W min respectively represent the window setting coefficients of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times; w 01 and w 02 respectively represent the weight values of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times;

[0037] When there is a situation where the window setting coefficient corresponding to each of the dynamic feature data is not lower than the preset window setting coefficient, then the window setting coefficient corresponding to each of the dynamic feature data is used to obtain the corresponding duration of the rolling window;

[0038] Among them, the corresponding duration of the rolling window is obtained through the following formula:

[0039]

[0040] Among them, T w represents the corresponding duration of the rolling window; n represents the number of dynamic feature data whose window setting coefficients are not lower than the preset window setting coefficient; m represents the number of dynamic feature data whose window setting coefficients are lower than the preset window setting coefficient; W iDenote the window setting coefficient of the dynamic feature data where the i-th window setting coefficient is not lower than the preset window setting coefficient; T i Denote the time scale of the dynamic feature data where the i-th window setting coefficient is not lower than the preset window setting coefficient; W j Denote the window setting coefficient of the dynamic feature data where the j-th window setting coefficient is lower than the preset window setting coefficient. W th Denote the preset window setting coefficient;

[0041] Preferably, the construction of the additional model in S2 specifically includes the following steps:

[0042] S21: Collect the professional capabilities of the attending physician himself, past treatment cases, and the basic information and performance data of the caregivers respectively;

[0043] S22: Convert the above data into quantifiable features, perform numerical processing on the data that can be numerically counted, perform categorical variable coding on some data, and finally conduct a comprehensive score. According to expert opinions or historical data, assign weights to each feature, and calculate the comprehensive experience scores of the doctor and the caregiver;

[0044] S23: Select a model for training, and finally obtain a doctor experience model and a caregiver ability model. Assign weights according to the historical performance indicators of the two models. Let the output of the doctor experience model be S doctor , and the output of the caregiver ability model be S nurse , and the output S of the final additional model additional , can be calculated by the following formula:

[0045] S additional = w doctor ·S doctor + w nurse ·S nurse , where: w doctor and w nurse are the weights of the doctor experience model and the caregiver ability model respectively, and satisfy the constraint condition: w doctor + w nurse = 1.

[0046] As new data accumulates, regularly re-evaluate the performance of the two models and update the weights. Use indicators to evaluate the overall performance of the additional model, compare the effects of using the doctor experience model or the caregiver ability model alone, and verify the effectiveness of the weighted fusion method.

[0047] Preferably, the regular re-evaluation of the performance of the two models and the update of the weights specifically include the following steps:

[0048] S231: Obtain the latest data to reflect the current situation, collect new data on postoperative risk cases, including the postoperative recovery of patients, the experience scores of attending physicians, and the ability scores of caregivers, ensure the quality of the new data, and perform cleaning, missing value handling, and standardization;

[0049] S232: Evaluate the performance of the doctor experience model and the caregiver ability model on the new data. Divide the new data into a training set and a test set, use the existing doctor experience model and caregiver ability model to predict the test set, calculate the metrics, compare the performance metrics on the new data with the performance on the historical data, and determine whether the model needs to be adjusted;

[0050] S233: Dynamically adjust the weights according to the performance metrics of the model. Calculate the relative importance of the two models based on the AUC or F1 score on the new data. If the performance of a certain model drops significantly, such as the AUC being lower than the set threshold, its weight can be reduced or the model can be retrained.

[0051] Preferably, the integration of the S3 main model and the additional model specifically includes the following steps:

[0052] S31: Collect the output data of the main model and the additional model. For each patient, record the output S of the main model main and the output S of the additional model additional ;

[0053] S32: Construct the input features for non - linear fusion, X = [S main , S additional ;

[0054] S33: Select a neural network for fusion. Construct a multi - layer perceptron (MLP) model. The input is the output of the main model and the additional model, and the output is the final postoperative risk prediction value. The network structure can include several hidden layers, and use ReLU or Sigmoid as the activation function;

[0055] S34: Use the historical data to train the non - linear fusion model. Divide the data into a training set (70% - 80%) and a test set (20% - 30%). Use the training set to train the selected non - linear model, use K - fold cross - validation to optimize the model parameters, and adjust the hyperparameters to improve the model performance. Hyperparameters such as the number of layers of the neural network, the learning rate, the number of trees and depth of XGBoost, etc.;

[0056] S35: Evaluate the performance of the non - linear fusion model and generate the final postoperative risk prediction value.

[0057] Another technical problem to be solved by the present invention is to provide a system for constructing a coronary intervention postoperative risk prediction model. The present system includes the following modules:

[0058] Data acquisition module: Responsible for collecting clinical data of patients from channels such as hospital information system (HIS), electronic medical record system (EMR), and questionnaires, including but not limited to age, gender, medical history, laboratory test results, imaging data, surgical records, medication records, etc. It uses wearable devices to monitor patients' physiological indicators in real time and obtains recent life characteristic data of patients, including but not limited to patients' diet, work and rest time, sleep quality, exercise and physical activities, mental health and emotional state, health awareness and behavior, environmental exposure, work type, social and economic factors, culture and beliefs, and digital behavior. It retrieves doctors' treatment experience data and caregiver information by checking HIS and questionnaires;

[0059] Data processing module: Performs preprocessing operations such as cleaning, encoding, and standardization on the collected data;

[0060] Main model training module: Implements the training, validation, and optimization functions of the main model;

[0061] Additional model training module: Respectively implements the training of the doctor experience model and the caregiver model;

[0062] Model integration module: Fuses the results of the main model and the additional models to generate the final postoperative risk prediction value;

[0063] User interaction module: Provides a graphical interface that allows doctors to input patient-related information and view the prediction results.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] A method and system for constructing a coronary intervention postoperative risk prediction model proposed by the present invention combines a main model and additional models. The main model is based on static and dynamic characteristic data of patients and comprehensively models through a regression model and a time series model to capture the individualized health status changes of patients. The additional models evaluate the impact of the medical team and relative care on postoperative risk by quantifying doctors' experience and caregivers' capabilities, making up for the deficiencies of traditional models that only rely on patient data. A non-linear model such as a neural network is used to fuse the results of the main model and the additional models, which can effectively capture the complex interaction relationship between the two, significantly improve the prediction accuracy, and ensure high data quality and comparability through operations such as missing value processing, outlier detection, encoding, and normalization, reducing the impact of noise on the model. Cross-validation and hyperparameter tuning techniques are used to optimize the model performance and ensure the stable performance of the model on unseen data. As new data accumulates, the performance of the doctor experience model and the caregiver capability model is regularly re-evaluated, and the weights are dynamically adjusted to keep the model in the best state at all times. Description of the Drawings

[0066] Figure 1 This is a module diagram of the system for constructing a risk prediction model after coronary intervention of the present invention. Specific implementation manners

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0068] To solve the problem in the prior art that the hospital physical examination information of patients is used as all the basic data for model construction, and the data is all based on the patients as the entire risk subjects, excluding the postoperative risks of the medical staff and caregivers around the patients, please refer to Figure 1 , the following technical solutions are provided in this embodiment:

[0069] A method for constructing a risk prediction model after coronary intervention includes the following steps:

[0070] S1. Construct a main model: Obtain the clinical data and life characteristic data of patients, process the data, divide the static characteristic data and dynamic characteristic data. The criteria for dividing the static characteristic data and dynamic characteristic data are time scale and acquisition frequency. The static characteristic data represents data that remains stable within a period of time and will not change significantly in the short term, usually reflecting the basic attributes or long-term health status of patients, including but not limited to the personal information of patients, such as age, gender; unchangeable physiological indicators, such as past medical history, genetic factors; long-term habits in lifestyle, such as eating habits, smoking history. The dynamic characteristic data represents data that fluctuates over time, usually reflecting the health status or behavior pattern of patients within a specific time period. For example: real-time monitoring data, such as heart rate, blood pressure, blood glucose level; short-term behavior or environmental changes, such as recent exercise frequency, psychological stress level;

[0071] Based on the static characteristic data, complete the construction of a regression model through a machine learning algorithm, divide the training set and the test set, optimize the model parameters using cross-validation, construct a time series model using the dynamic characteristic data, splice the features extracted by the time series model with the static characteristic data into a new feature vector, retrain the regression model using the comprehensive feature vector, evaluate the model performance, and evaluate the model performance using indicators such as accuracy, recall rate, F1 score, and area under the ROC curve AUC;

[0072] S11: Collect the static characteristic data of patients, process missing values, detect and process outliers, encode non-numerical data, and normalize numerical data to ensure that different features have the same dimension;

[0073] S12: Calculate the correlation coefficient between each feature and the postoperative risk, preliminarily screen the features with relatively high correlation, use the LASSO regression algorithm to automatically screen important features through L1 regularization, and use the random forest algorithm to select key features according to the importance scores of the features;

[0074] S13: Divide the training set according to the ratio of 70%-80%, and use the remaining part as the test set. Use stratified sampling to ensure the consistent distribution of the training set and the test set;

[0075] S14: Select machine learning algorithms to construct a regression model. Common algorithms include logistic regression, random forest, support vector machine SVM, XGBoost, etc. Use the training set to train the selected algorithm, use K-fold cross-validation to optimize the model parameters, where K-fold is such as 5-fold or 10-fold, adjust the hyperparameters to improve the model performance, evaluate the performance of the regression model. The evaluation metrics include Accuracy: the proportion of correct predictions, Recall: the proportion of true positives correctly identified, F1 score: the harmonic mean of accuracy and recall, Area Under the ROC Curve AUC: an indicator to measure the discrimination ability of the model. Compare the performances of different algorithms and select the optimal model;

[0076] S15: Process the dynamic feature data, convert the format of the dynamic feature data into a two-dimensional array, where each row represents a time point and each column represents a variable. For missing values, interpolation methods (such as linear interpolation, spline interpolation) can be used to fill them. Use rolling window statistics (such as moving average, standard deviation) to detect outliers, correct or delete the outliers, and normalize or standardize the numerical data to ensure that different features have the same dimension.

[0077] Select a time series model, classical time series models:

[0078] ARIMA: Suitable for stationary time series data.

[0079] SARIMA: Suitable for time series data with seasonal characteristics.

[0080] Modern deep learning models:

[0081] LSTM: Good at dealing with long-term dependencies.

[0082] GRU: Similar to LSTM, but with a simpler structure and higher computational efficiency.

[0083] TCN: A time series model based on convolutional neural network.

[0084] Specifically, the process of setting the corresponding duration of the rolling window includes:

[0085] Extract the time scale corresponding to each dynamic feature data;

[0086] Obtain the time scale difference between every two dynamic feature data by using the time scale corresponding to each dynamic feature data;

[0087] Compare the time scale difference between the two dynamic feature data with a preset time scale difference threshold;

[0088] When the time scale difference between every two dynamic feature data included in all dynamic feature data is less than the preset time scale difference threshold, use the maximum value of the time scale of the corresponding dynamic feature data in all dynamic feature data as the corresponding duration of the rolling window;

[0089] When there is a situation where the time scale difference between every two dynamic feature data included in all dynamic feature data is not less than the preset time scale difference threshold, set the corresponding duration of the rolling window by using the time scale corresponding to each dynamic feature data and the number of data acquisitions within the time scale for each dynamic feature data.

[0090] The technical effect of the above technical solution is as follows: This solution can automatically adjust the duration of the rolling window according to the time scale of the dynamic feature data. This means that for different data sets or different features, the length of the rolling window can vary flexibly to adapt to the actual characteristics of the data. This adaptability helps to improve the accuracy and efficiency of data analysis. By comparing the time scale difference between every two dynamic feature data with the preset time scale difference threshold, this solution can identify the stability or variability of the data in terms of time scale. When the time scales of all data are relatively close (i.e., the difference is less than the threshold), selecting the maximum time scale as the rolling window duration can ensure that the window covers a long enough time period to capture the main trend of the data. When the time scale differences of the data are large, setting the rolling window duration by combining the time scale and the number of data acquisitions can more precisely balance the temporal characteristics and sampling density of the data. This solution avoids the problems of overfitting (window too small) or information loss (window too large) that may be caused by a fixed window by intelligently selecting the rolling window duration. This helps to improve the efficiency and effect of data processing, especially when dealing with large-scale or high-frequency data sets. The threshold setting in the solution allows for adjustment according to actual needs, increasing flexibility. At the same time, by considering the time scale and the number of data acquisitions of each dynamic feature data, this solution has good robustness to outliers or irregular sampling in the data and can handle various data situations more robustly.

[0091] In summary, this technical solution improves the accuracy, efficiency, and flexibility of data analysis by dynamically adjusting the length of the rolling window to adapt to the characteristics of different data, while maintaining good robustness and easy implementation.

[0092] Specifically, the corresponding duration of the rolling window is set by using the time scale corresponding to each dynamic feature data and the number of data collection times of each dynamic feature data within the time scale, including:

[0093] The time scale for extracting each dynamic feature data;

[0094] Extract the data collection times of each dynamic feature data within the time scale;

[0095] The window setting coefficient corresponding to each dynamic feature data is obtained by using the time scale corresponding to each dynamic feature data and the number of data collection times of each dynamic feature data within the time scale;

[0096] The window setting coefficient corresponding to each dynamic feature data is obtained by the following formula:

[0097]

[0098] Wherein, W(T, N) represents the window setting coefficient corresponding to each dynamic feature data. The larger the W(T, N) value is, the more drastic the dynamic change of the data is (or the more complex the characteristics are), and a more flexible window setting is required (such as shortening the window length or increasing the update frequency); the smaller the W(T, N) value is, the weaker the data dynamics is or the sampling is sparse, and the current window length should be appropriately increased; N represents the number of data collection times for each dynamic feature data within the time scale; T represents the time scale of each dynamic feature data; α represents the dynamic sensitivity factor, which is used to adjust the sensitivity of the collection corresponding to the change of dynamic feature data. The value range is 0.7-1.2, preferably 1.0; β represents the preset baseline sampling rate, which is used to represent the standard number of samplings per unit time; r represents the time decay constant, with a value range of 0.3-0.7, preferably 0.5, which is used to control the attenuation effect of the time scale on the window; T0 represents the preset reference time scale, which is used to normalize the time dimension; k represents the nonlinear correction index, which is used to adjust the nonlinear influence of the time scale, with a value range of 1.5-2.3, preferably 2; the above technical solution automatically adapts to the second-level to hour-level data through the modulation term to avoid manual segmentation settings. The logarithmic term suppresses high-frequency sampling interference, and the exponential term filters invalid time scales. The parameter open interface supports integration with the AIoT platform to meet the dynamic optimization needs of Industry 4.0. This formula significantly surpasses the traditional linear model in terms of functional coupling and parameter interpretability, and meets the requirements of patent law for creativity.

[0099] Compare the window setting coefficient corresponding to each piece of dynamic feature data with a preset window setting coefficient threshold;

[0100] When the window setting coefficient corresponding to each piece of dynamic feature data is lower than the preset window setting coefficient, then obtain the corresponding duration of the rolling window by combining the time scales of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times with the window setting coefficient corresponding to the dynamic feature data;

[0101] Among them, the corresponding duration of the rolling window is obtained through the following formula:

[0102]

[0103] Among them, T k represents the corresponding duration of the rolling window; T max and T min respectively represent the time scales of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times; W max and W min respectively represent the window setting coefficients of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times; w 01 and w 02 respectively represent the weight values of the dynamic feature data corresponding to the maximum data acquisition times and the minimum data acquisition times;

[0104] When there is a situation where the window setting coefficient corresponding to each piece of dynamic feature data is not lower than the preset window setting coefficient, then obtain the corresponding duration of the rolling window by the window setting coefficient corresponding to each piece of dynamic feature data;

[0105] Among them, the corresponding duration of the rolling window is obtained through the following formula:

[0106]

[0107] Among them, T w represents the corresponding duration of the rolling window; n represents the number of pieces of dynamic feature data whose window setting coefficients are not lower than the preset window setting coefficient; m represents the number of pieces of dynamic feature data whose window setting coefficients are lower than the preset window setting coefficient; W i represents the window setting coefficient of the i-th piece of dynamic feature data whose window setting coefficient is not lower than the preset window setting coefficient; T i represents the time scale of the i-th piece of dynamic feature data whose window setting coefficient is not lower than the preset window setting coefficient; W j represents the window setting coefficient of the j-th piece of dynamic feature data whose window setting coefficient is lower than the preset window setting coefficient. W th represents the preset window setting coefficient;

[0108] The technical effects of the above technical solution are as follows: By quantifying the data acquisition density, compressing the marginal benefit of high sampling rate through natural logarithm, introducing a logarithmic function to avoid numerical explosion caused by linear growth, and at the same time retaining the sensitivity to density differences. Suppressing the window setting weights for ultra-long or ultra-short time scales, preventing extreme values from interfering, combining an exponential decay function with a power-law correction to strengthen the penalty for unconventional time scales;

[0109] w 01 and w 02 respectively represent the weight values of the dynamic feature data corresponding to the maximum number of data acquisition times and the minimum number of data acquisition times; W max and W min respectively represent the window setting coefficients of the dynamic feature data corresponding to the maximum number of data acquisition times and the minimum number of data acquisition times; and are used to adjust the weight according to the window setting coefficient. The window setting coefficient reflects a certain characteristic or requirement of data acquisition. Through this division operation, the role of the weight is adapted to the window setting situation. T max and T min respectively represent the time scales of the dynamic feature data corresponding to the maximum number of data acquisition times and the minimum number of data acquisition times; and multiply the adjusted weights by the corresponding time scales respectively to obtain two values related to the weight and the time scale. This step combines the time scale with the weight adjusted by the window setting coefficient to reflect the influence of the time characteristics of data at different acquisition times on the duration of the rolling window. After adding the above two product results and then multiplying by 0.5, that is finally obtain the corresponding duration T of the rolling window kThe operation of taking the average comprehensively considers the data situations under two different acquisition times (the maximum and the minimum), balances their contributions to the rolling window duration, and thus obtains a relatively reasonable rolling window duration value. The technical content mentions that this formula is used when the window setting coefficients corresponding to each dynamic feature data are all lower than the preset value. The formula comprehensively considers multiple factors such as the data acquisition times (reflected by the maximum and the minimum), the time scale, the window setting coefficients, and the weights. Different acquisition times may reflect the data acquisition situations of the system in different states. The time scale describes the time characteristics of the data. The window setting coefficients reflect the setting requirements of data acquisition. The weights can highlight the role of key data. By organically combining these factors, the formula can more comprehensively determine the rolling window duration according to the actual data acquisition and feature situations, which conforms to the idea of comprehensive analysis in the technology. In the actual acquisition of dynamic feature data, the data acquisition times and window setting situations may vary. The formula can adapt to the differences in data characteristics by assigning different weights to the data under different acquisition times and making adjustments in combination with the window setting coefficients. For example, for data with a large number of acquisitions but a small window setting coefficient (which may mean that the data is denser and more important), through the operations of weights and window setting coefficients, it can be given appropriate influence in the calculation of the rolling window duration, so that the calculated duration can more accurately reflect the actual value and application requirements of the data. From the perspective of technical requirements, it is necessary to obtain the corresponding duration of the rolling window according to specific conditions (the window setting coefficient is lower than the preset value). The calculation process of the formula first makes adjustments related to the window setting coefficients for the weights, then calculates in combination with the time scale, and finally takes the average to obtain the result. This process has a clear logic, is in line with the goal of determining the rolling window duration in the technology, and can reasonably calculate the corresponding duration of the rolling window that conforms to the actual situation according to the established rules and requirements, with technical rationality.

[0110] Meanwhile, this solution dynamically adjusts the duration of the rolling window by introducing multiple parameters (such as the dynamic sensitivity factor α, the reference sampling rate β, the time decay constant r, the reference time scale T0, and the non-linear correction exponent k), enabling the window time scale setting to highly adapt to datasets with different characteristics. Through the calculation of the window setting coefficient W(T, N), this solution can automatically identify the degree of dynamic change and sampling density of the data, thereby flexibly adjusting the window length to meet the requirements in different scenarios. By quantifying the data acquisition density and introducing the logarithmic function, this solution can effectively suppress the interference caused by high-frequency sampling while retaining the sensitivity to density differences, thus improving the accuracy of data processing. The exponential decay function combined with the power-law correction strengthens the penalty for unconventional time scales, helping to prevent the interference of extreme values on the window setting and further enhancing the stability of data processing. The formulas and parameter designs in this solution have clear physical meanings and interpretability, enabling technicians to more easily understand and adjust the model. The parameter open interface supports integration with the AIoT platform, meets the dynamic optimization requirements of Industry 4.0, realizes a high degree of functional coupling, and facilitates deployment and optimization in practical applications. By automatically adapting the modulation term to data from seconds to hours, this solution can automatically adjust the window duration, avoiding the cumbersome process of manual segmentation setting in traditional methods. When the data has strong dynamics or a high sampling density, the window setting coefficient is large, and the window duration will be correspondingly shortened or the update frequency will increase to adapt to the rapid changes in the data. When the data has weak dynamics or a low sampling density, the window setting coefficient is small, and the window duration will be appropriately increased to capture the long-term trend of the data. This solution has good robustness to outliers or irregular sampling in the data and can intelligently adjust the window setting to adapt to various complex data situations. The basic framework and parameter design of this solution are scalable and can be easily extended to more complex scenarios or integrated with other systems.

[0111] For n dynamic feature data with window setting coefficients not lower than the preset value W th Multiply the window setting coefficient Wi of each data by its time scale Ti and then sum them. This step comprehensively considers the window setting situation and time characteristics of this type of data. Through the method of product summation, a total value related to these data is obtained, reflecting their comprehensive contribution to the duration of the rolling window. For n dynamic feature data with window setting coefficients not lower than the preset value W th Sum the window setting coefficients Wi. It is to perform a weighted average on the data with window setting coefficients not lower than the preset value according to the window setting coefficients to obtain an average reference value based on the time scale of this type of data, which serves as the basic part for calculating the duration of the rolling window.

[0112] The window setting coefficient for the window is lower than the preset value W th for the m dynamic feature data with a window setting coefficient W j are summed up, reflecting the total situation of the window setting coefficients of this part of the data. By taking the ratio with m Wmax, the total amount of the window setting coefficients of the data lower than the preset value is scaled and adjusted, considering the influence of the number of data and the maximum window setting coefficient, to obtain an adjustment factor for measuring the additional influence degree of this part of the data on the rolling window duration. The overall formula for this part is +1 to form a comprehensive adjustment coefficient for correcting the average reference value obtained in the first part.

[0113] In the technical content, according to the size relationship between the window setting coefficient and the preset value, the dynamic feature data are clearly divided into two categories (not lower than the preset value and lower than the preset value). The formula also processes these two categories of data separately during the calculation process. The first part focuses on the data with a window setting coefficient not lower than the preset value, and determines the basic time scale reference value through weighted average; the second part analyzes the data with a window setting coefficient lower than the preset value to obtain an adjustment factor for correcting the basic reference value. This way of distinguishing and processing is consistent with the idea of classifying data in the technical content, and can reasonably calculate the rolling window duration according to the characteristics of different data. The formula comprehensively considers multiple factors such as the number of data, the window setting coefficient, and the time scale. Different numbers of data reflect the quantitative differences between the two categories of data, the window setting coefficient reflects the characteristics of data acquisition, and the time scale describes the time attributes of the data. By organically combining these factors, the formula can comprehensively determine the rolling window duration according to the actual data situation, meeting the requirements of comprehensively analyzing data to determine the window duration in the technology. From the perspective of technical requirements, it is necessary to calculate the corresponding duration of the rolling window according to the situation of the window setting coefficient. The calculation process of the formula first processes the data in different situations respectively to obtain the basic reference value and the adjustment coefficient, and then multiplies them to obtain the final result. The entire process is logically coherent and can reasonably calculate the rolling window duration that meets the actual application scenario according to the rules set by the technology and the characteristics of the data, with technical rationality.

[0114] In summary, through features such as a highly adaptive window setting mechanism, improving the accuracy and efficiency of data processing, parameter interpretability and functional coupling, avoiding the cumbersome manual segment setting, intelligently balancing data dynamics and sampling density, and strong robustness, this technical solution significantly improves the ability of data processing and analysis, meeting the dynamic optimization requirements in complex application scenarios such as Industry 4.0.

[0115] Perform model training, input time series data (such as the time series of postoperative heart rate and blood pressure), output the predicted values of the time series model (such as the blood pressure level or abnormal detection result at a certain future moment), extract the key features of the time series model, for example: dynamic trends, such as whether the blood pressure continues to rise; volatility, such as the standard deviation of the heart rate; abnormal detection results, such as whether there are abnormal fluctuations, evaluate the performance of the time series model, and the evaluation metrics include mean squared error (MSE): measuring the difference between the predicted value and the true value; mean absolute error (MAE): measuring the average deviation between the predicted value and the true value; R 2 score: measuring the ability of the model to explain the variability of the data, comparing the performance of different models, and selecting the optimal model;

[0116] S16: Concatenate the dynamic features (such as the trend and volatility of the heart rate) extracted by the time series model with the static feature data, and use a regression model to model this feature vector to generate the final postoperative risk predicted value;

[0117] S2. Build an additional model: Build a scoring system for the experience of attending physicians and a scoring system for the capabilities of caregivers, predict the postoperative risk of patients based on the doctor's experience and the caregiver's ability score, and weight and fuse the results of the doctor experience model and the caregiver model to form the output of the final additional model;

[0118] S21: Collect the professional capabilities, past treatment cases of attending physicians and the basic information and performance data of caregivers respectively. The data of attending physicians include but are not limited to: basic information of the doctor: such as age, gender, years of practice;

[0119] Professional background of the doctor: such as educational level (master, doctor), professional title (attending physician, deputy chief physician, chief physician);

[0120] Clinical experience of the doctor: such as the success rate of coronary intervention surgery, the number of surgeries completed per year, and the ability to handle complex cases;

[0121] Training experience of the doctor: such as whether they have participated in advanced training or international conferences in related fields;

[0122] The basic information of caregivers includes but is not limited to: basic information of caregivers: such as age, gender, mental health and emotional state, and the work they are engaged in;

[0123] Professional skills of caregivers: health awareness and behavior, social and economic factors, and whether they have participated in professional training related to postoperative care;

[0124] Work performance of caregivers: such as the relationship with patients, daily care or accompanying time;

[0125] S22: Convert the above data into quantifiable features. For data that can be numerically counted, perform numerical processing. For example, for the years of practice: record it directly in years; for the surgical success rate: calculate the proportion of successful cases in historical surgeries; for the annual number of surgeries: count the number of coronary intervention surgeries completed by the doctor each year. Encode some data as categorical variables. For example, for professional title: use one-hot encoding or label encoding; for educational level: assign values according to the educational attainment, such as undergraduate = 1, master = 2, doctor = 3. Finally, perform a comprehensive score. According to expert opinions or historical data, assign weights to each feature and calculate the comprehensive experience scores of doctors and caregivers.

[0126] S23: Select a model for training to finally obtain a doctor experience model and a caregiver ability model. Assign weights according to the historical performance metrics (such as AUC, F1 score) of the two models. Let the output of the doctor experience model be S doctor , and the output of the caregiver ability model be S nurse , and the output S of the final additional model additional , can be calculated by the following formula:

[0127] S additional = w doctor ·S doctor + w nurse ·S nurse , where: w doctor and w nurse are the weights of the doctor experience model and the caregiver ability model respectively, and satisfy the constraint condition: w doctor + w nurse = 1.

[0128] As new data accumulates, regularly re-evaluate the performance of the two models and update the weights. Use metrics to evaluate the overall performance of the additional model, compare the effects of using the doctor experience model or the caregiver ability model alone, and verify the effectiveness of the weighted fusion method.

[0129] Regularly re-evaluating the performance of the two models and updating the weights specifically includes the following steps:

[0130] S231: Obtain the latest data to reflect the current situation. Collect new postoperative risk case data, including the postoperative recovery of patients, the experience scores of attending doctors, and the ability scores of caregivers. Ensure the quality of the new data, and perform cleaning, missing value handling, and standardization.

[0131] S232: Evaluate the performance of the doctor experience model and the caregiver ability model on new data. Divide the new data into a training set and a test set. Use the existing doctor experience model and caregiver ability model to make predictions on the test set, calculate the metrics, compare the performance metrics on the new data with those on the historical data, and determine whether the models need to be adjusted;

[0132] S233: Dynamically adjust the weights according to the performance metrics of the models. Calculate the relative importance of the two models based on the AUC or F1 score on the new data. For example, assume the AUC of the doctor experience model is 0.85 and the AUC of the caregiver ability model is 0.75. Then the weights can be allocated proportionally:

[0133] w doctor = 0.85 / (0.85 + 0.75) = 0.53, w nurse = 0.75 / (0.85 + 0.75) = 0.47

[0134] If the performance of a certain model drops significantly, such as the AUC being lower than the set threshold, its weight can be reduced or the model can be retrained.

[0135] S3: Integration of the main model and the additional model: Non-linearly fuse the output results of the main model and the additional model to generate the final postoperative risk prediction value.

[0136] S31: Collect the output data of the main model and the additional model. For each patient, record the output S main of the main model and the output S additional of the additional model;

[0137] S32: Construct the input features for non-linear fusion, X = [S main , S additional ;

[0138] S33: Select a neural network for fusion. Build a multi-layer perceptron (MLP) model with the outputs of the main model and the additional model as the input and the final postoperative risk prediction value as the output. The network structure can include several hidden layers, and use ReLU or Sigmoid as the activation function;

[0139] S34: Use the historical data to train the non-linear fusion model. Divide the data into a training set (70% - 80%) and a test set (20% - 30%). Use the training set to train the selected non-linear model, use K-fold cross-validation to optimize the model parameters, and adjust the hyperparameters to improve the model performance. The hyperparameters include the number of layers of the neural network, the learning rate, the number of trees and depth of XGBoost, etc.;

[0140] S35: Evaluate the performance of the non-linear fusion model and generate the final postoperative risk prediction value.

[0141] The following data is known:

[0142] Patient number <![CDATA[S main > <![CDATA[S additional > True postoperative risk 1 0.8 0.7 1 2 0.6 0.9 0 3 0.9 0.8 1

[0143] The input feature vector is:

[0144] X = [S main , S additional

[0145] Patient 1: X1 = [0.8, 0.7]

[0146] Patient 2: X2 = [0.6, 0.9]

[0147] Patient 3: X3 = [0.9, 0.8]

[0148] Use a neural network model:

[0149] Input layer: 2 nodes, corresponding to S main and S additional .

[0150] Hidden layer: 1 layer, containing 10 nodes, with the activation function being ReLU.

[0151] Output layer: 1 node, with the activation function being Sigmoid, and the output range being [0, 1].

[0152] Use the training set to train the neural network, and adjust the learning rate and the number of iterations to optimize the model performance. For a new patient, assume the output of the main model S main = 0.75, and the output of the additional model S additional = 0.85.

[0153] Input into the non - linear fusion model to obtain the final prediction value:

[0154] S final = 0.88.

[0155] Another technical problem to be solved by the present invention is to provide a system for constructing a risk prediction model after coronary intervention. This system includes the following modules:

[0156] ​Data collection module: Responsible for collecting patients' clinical data from channels such as the hospital information system (HIS), electronic medical record system (EMR), and questionnaires, including but not limited to age, gender, medical history, laboratory test results, imaging data, surgical records, medication records, etc. It uses wearable devices to real-time monitor patients' physiological indicators such as heart rate, blood pressure, blood sugar, etc., and obtains patients' recent life characteristic data, including but not limited to patients' diet, work and rest time, sleep quality, exercise and physical activities, mental health and emotional state, health awareness and behavior, environmental exposure, work type, social and economic factors, culture and beliefs, and digital behavior, etc. It retrieves doctors' treatment experience data and caregiver information by checking HIS and questionnaires;

[0157] The collection tools for life characteristic data include:

[0158] Wearable devices and smart devices, such as: smart bracelets / watches, which collect data such as heart rate, sleep quality, steps, activity volume, blood oxygen saturation, etc.; smart weighing scales, which collect data such as weight, body fat percentage, muscle mass, etc.; smart blood pressure monitors, which collect blood pressure fluctuations;

[0159] Mobile applications, such as: diet record Apps, which collect daily diet types, intake, and nutritional components such as calories, fat, sugar, salt, etc.; sleep monitoring Apps, which collect data such as sleep duration, deep sleep time, and number of night awakenings; life habit record Apps, which collect smoking, drinking, exercise habits, etc.;

[0160] Questionnaires and electronic diaries, such as: structured questionnaires, which collect work types, work pressure, living environment, personality traits such as type A personality, anxiety tendency; electronic diaries, which collect daily mood changes, stress events, social activities, etc.;

[0161] Environmental sensors, such as: air quality monitors, which collect pollutant concentrations such as PM 2.5 , NO2, CO, etc.; noise monitors, which collect environmental noise levels;

[0162] Social media and behavior data analysis, such as: social media data, which collect patients' posts, emotional tendencies, and social interaction frequencies; behavior data, which collect mobile phone usage duration, exercise App usage frequencies, etc.;

[0163] Data processing module: Performs preprocessing operations on the collected data, such as cleaning, encoding, standardization, etc.;

[0164] Main model training module: Implement the training, validation, and optimization functions of the main model, screen out the features that have a significant impact on postoperative risk prediction, select a machine learning algorithm suitable for the main model, use the training set to train the selected algorithm, evaluate the performance of the model on unseen data, improve the model performance by adjusting hyperparameters, and evaluate the overall performance of the main model;

[0165] Additional model training module: Implement the training of the doctor experience model and the caregiver model respectively;

[0166] Model integration module: Integrate the results of the main model and the additional models, determine the weights of the main model and the additional models, fuse the results of the main model and the additional models to generate the final prediction value, and generate the final postoperative risk prediction value;

[0167] User interaction module: Provide a graphical interface that allows doctors to input patient-related information and view the prediction results.

[0168] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0169] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

Claims

1. A method for constructing a risk prediction model after coronary intervention, characterized in that: The following steps are involved: The clinical data and life characteristic data of patients are processed, and static characteristic data and dynamic characteristic data are divided. The dynamic characteristic data are used to build a time series model, and the features extracted by the time series model are spliced ​​with the static characteristic data into a new feature vector to build the main model; the processing of dynamic characteristic data includes converting the data format into a two-dimensional array, using interpolation to fill missing values, and detecting and correcting abnormal values ​​through rolling window statistics. The rolling window duration is dynamically set according to the time scale of the dynamic characteristic data and the number of acquisitions; Construct the attending physician experience scoring system and the caregiver ability scoring system, weightedly fuse the results of the physician experience model and the caregiver model to form the final additional model output, and nonlinearly fuse the output results of the main model and the additional model to generate the final postoperative risk prediction value.

2. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 1, characterized in that: The criteria for dividing the static feature data and the dynamic feature data are the time scale and the acquisition frequency. The static feature data indicates that the data remain stable over a period of time, while the dynamic feature data indicates that the data fluctuates over time.

3. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 1, characterized in that: Building the main model specifically includes the following steps: Collect static feature data of patients, handle missing values, detect and process outliers, encode non-numeric data, normalize numerical data, and ensure that different features have the same dimension; The correlation coefficient between each feature and postoperative risk was calculated, and the features with high correlation were preliminarily screened. The LASSO regression algorithm was used to automatically screen important features through L1 regularization, and the random forest algorithm was used to select key features according to the feature importance score; The training set is divided into 70%-80% and the rest is used as the test set. Stratified sampling is used to ensure that the distribution of the training set and the test set is consistent. Select a machine learning algorithm to build a regression model, use the training set to train the selected algorithm, use K-fold cross-validation to optimize model parameters, adjust hyperparameters to improve model performance, evaluate the performance of the regression model, and evaluate indicators including accuracy, recall, F1 score, and area under the ROC curve (AUC). Compare the performance of different algorithms and select the optimal model; Process dynamic feature data, select a time series model, perform model training, input time series data, output the predicted value of the time series model, extract the key features of the time series model, evaluate the performance of the time series model, compare the performance of different models, and select the optimal model; The dynamic features extracted by the time series model are spliced ​​with the static feature data, and the feature vector is modeled using a regression model to generate the final postoperative risk prediction value.

4. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 3, characterized in that: The processing method of the dynamic feature data is as follows: the format of the dynamic feature data is converted into a two-dimensional array, in which each row represents a time point and each column represents a variable. For missing values, interpolation is used to fill them, and rolling window statistics are used to detect outliers, and outliers are corrected or deleted. The numerical data is normalized or standardized to ensure that different features have the same dimension.

5. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 4, characterized in that: The process of setting the duration of the rolling window includes: Extract the time scale corresponding to each dynamic feature data; Using the time scale corresponding to each dynamic feature data, obtaining the time scale difference between every two dynamic feature data; Comparing the time scale difference of the two dynamic feature data with a preset time scale difference threshold; When the time scale difference between every two dynamic feature data included in all dynamic feature data is less than the preset time scale difference threshold, the maximum time scale value of the corresponding dynamic feature data in all dynamic feature data is used as the corresponding time length of the rolling window; When the time scale difference between every two dynamic feature data contained in all dynamic feature data is not less than the preset time scale difference threshold, the corresponding time length of the rolling window is set by using the time scale corresponding to each dynamic feature data and the number of data collection times of each dynamic feature data within the time scale.

6. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 5, characterized in that: The corresponding duration of the rolling window is set by using the time scale corresponding to each dynamic feature data and the number of data collection times within the time scale for each dynamic feature data, including: The time scale for extracting each dynamic feature data; Extract the data collection times of each dynamic feature data within the time scale; The window setting coefficient corresponding to each dynamic feature data is obtained by using the time scale corresponding to each dynamic feature data and the number of data collection times of each dynamic feature data within the time scale; Comparing the window setting coefficient corresponding to each dynamic feature data with a preset window setting coefficient threshold; When the window setting coefficient corresponding to each dynamic feature data is lower than the preset window setting coefficient, the time scale of the dynamic feature data corresponding to the maximum data collection times and the minimum data collection times is combined with the window setting coefficient corresponding to the dynamic feature data to obtain the corresponding duration of the rolling window; When the window setting coefficient corresponding to each dynamic feature data is not lower than the preset window setting coefficient, the window setting coefficient corresponding to each dynamic feature data obtains the corresponding duration of the rolling window.

7. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 1, characterized in that: Building the additional model specifically includes the following steps: Collect the attending physician's own professional ability, past treatment cases, basic information of caregivers, and performance data; The acquired data is converted into quantitative features, the data that can be numerically counted is processed numerically, some data is encoded as categorical variables, and finally a comprehensive score is calculated. According to expert opinions or historical data, a weight is assigned to each feature, and the comprehensive experience score of doctors and caregivers is calculated; Select models for training, and finally obtain the doctor experience model and the caregiver capability model. Assign weights to the two models based on their historical performance indicators. As new data accumulates, regularly re-evaluate the performance of the two models and update the weights. Use indicators to evaluate the overall performance of the additional models, and compare the effects of using the doctor experience model or the caregiver capability model alone to verify the effectiveness of the weighted fusion method.

8. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 7, characterized in that: The periodic re-evaluation of the performance of the two models and updating the weights specifically includes the following steps: Obtain the latest data to reflect the current situation, collect new postoperative risk case data, including the patient's postoperative recovery, the attending physician's experience score, and the caregiver's ability score, ensure the quality of the new data, and perform cleaning, missing value processing, and standardization; Evaluate the performance of the doctor experience model and the caregiver competence model on the new data, divide the new data into a training set and a test set, use the existing doctor experience model and the caregiver competence model to predict the test set, calculate the indicators, compare the performance indicators on the new data with the performance on the historical data, and determine whether the model needs to be adjusted; Dynamically adjust weights based on the model's performance metrics, calculate the relative importance of the two models based on their AUC or F1 scores on new data, and reduce weights or retrain the model.

9. The method for constructing a risk prediction model after percutaneous coronary intervention according to claim 8, characterized in that: The integration of the main model and the additional model specifically includes the following steps: Collect the output data of the main model and the additional model. For each patient, record the output S of the main model. main and the output S of the additional model additional ; Construct input features for nonlinear fusion, X=[S main , S additional ]; Select neural networks for fusion and build a multi-layer perceptron model. The input is the output of the main model and the additional model. The output is the final postoperative risk prediction value. The network structure contains several hidden layers and uses ReLU or Sigmoid as the activation function. Use historical data to train nonlinear fusion models, divide the data into training sets and test sets, use the training set to train the selected nonlinear model, use K-fold cross-validation to optimize model parameters, and adjust hyperparameters to improve model performance; The performance of the nonlinear fusion model was evaluated to generate the final postoperative risk prediction value.

10. A system for constructing a risk prediction model after coronary intervention, used to implement the method for constructing a risk prediction model after coronary intervention according to any one of claims 1 to 9, characterized in that: include: The data collection module is responsible for collecting patients' clinical data from channels such as the hospital information system HIS, the electronic medical record system EMR, and questionnaires. It uses wearable devices to monitor patients' physiological indicators in real time, obtains patients' recent life characteristics data, and obtains doctors' treatment experience data and caregiver information by checking HIS and questionnaires. The data processing module cleans, encodes, and standardizes the collected data; Main model training module, which realizes the training, verification and optimization functions of the main model; Additional model training modules are provided to realize the training of doctor experience model and caregiver model respectively; The model integration module combines the results of the main model and the additional model to generate the final postoperative risk prediction value; The user interaction module provides a graphical interface that allows doctors to enter patient-related information and view prediction results.

Images