Medical safety risk assessment system and method
By designing a medical safety risk assessment system including control module, training module, prediction module and storage module, and using hybrid neural network model to perform risk assessment, the problem of lack of scientific, objective and quantifiable existing medical risk assessment methods is solved, real-time monitoring and prediction of medical risks is achieved, and the occurrence of medical disputes and adverse events is reduced.
Patent Information
- Application Number
- CN202411994302.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2025-04-29
AI Technical Summary
Existing medical risk assessment methods rely on medical personnel or expert systems for evaluation. They lack scientific, objective and quantifiable assessment methods, making it difficult to effectively identify and prevent medical disputes and adverse events.
Design a medical safety risk assessment system, including control module, training module, prediction module and storage module, and train and predict patients' medical indicators and complaints and adverse events through a hybrid neural network model (the combination of STAR-BiLSTM and XGBoost), to provide real-time risk assessment results.
Real-time monitoring and prediction of medical risks is achieved, helping hospitals identify high-risk patients in advance, reducing the occurrence of medical disputes and adverse events, and improving hospital service quality and patient satisfaction.
Smart Images
Figure CN119905222A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical risk assessment, and in particular to a medical safety risk assessment system and method. Background Art
[0002] In the modern medical system, hospitals are not only places for disease treatment, but also harbors for patients to seek health and receive care. As society pays more and more attention to the quality of medical services, hospitals implement the "patient-centered" service concept, requiring that patients' needs, preferences and values be given priority, improving patient satisfaction and the quality of medical services, reducing the number of patient complaints, and actively resolving existing disputes.
[0003] The present invention dynamically evaluates the patient's condition in real time, helps hospital managers proactively identify possible medical disputes, and can intervene in advance to minimize the occurrence of medical disputes. The medical risks defined in the present invention include patient complaint disputes and adverse medical events. Currently, early warnings for medical risks are mostly evaluated by medical personnel, or evaluated by scoring by an expert system composed of an expert group. Therefore, a scientific, objective, and quantifiable medical safety risk assessment method is proposed to provide an explainable basis for risk management of clinicians and medical departments, and to provide support for subsequent risk trend data statistics. Summary of the invention
[0004] In order to solve the technical problems in the above background, the present invention provides a medical safety risk assessment system, including a control module, a training module, a prediction module and a storage module;
[0005] The control module is used to control the overall workflow, including the scheduled start-up of the prediction module and the training module;
[0006] The training module is used to obtain the patient's medical indicators, complaints, and adverse events from the database, conduct model training, and store the training results;
[0007] The prediction module is used to obtain the model parameters from the storage module for loading, perform prediction operations, and send the prediction results to the database and the terminal through the network;
[0008] The storage module is used to store model parameters and log information and write data to the storage medium.
[0009] In the preferred solution, the training module trains the hybrid neural network model of STAR-BiLSTM and XGBoost, and judges it by the cost-sensitive error rate.
[0010] In the preferred solution, in the hybrid neural network model, several intermediate layers are designed between STAR-BiLSTM and XGBoost to further process the output of STAR-BiLSTM and then use it as the input of the XGBoost model.
[0011] In a preferred solution, the middle layer includes a summary statistics layer and an embedding layer;
[0012] Calculate statistics for the sliding window of time series data, including mean, median, maximum and minimum values, and input them into the XGBoost model as additional features;
[0013] The embedding layer is used to convert the sequence data output by STAR-BiLSTM into a dense vector to better capture the patterns in the sequence.
[0014] In a preferred solution, the training module also preprocesses the training data, and the preprocessing process specifically includes:
[0015] Time series data: For time series data, it is necessary to ensure that the timestamps of the data are aligned and handle missing values or outliers;
[0016] Feature engineering: extracting useful features from raw data, including time series indicators and static attributes;
[0017] Data splitting: Split the training data into training set, validation set and test set.
[0018] The present invention also provides a method for a medical safety risk assessment system, comprising the following steps:
[0019] S1. Initialize the timer, design the training cycle and prediction cycle according to the hospital's requirements, and set the timer's training task and prediction task start time;
[0020] S2, start the timer;
[0021] S3, the system timer queries whether the training task start time has arrived. If the training task start time has arrived, step 5 is executed; if the training task start time has not arrived, step S5 is executed, and then query whether the predicted task start time has arrived. If the predicted task start time has arrived, step S6 is executed;
[0022] S4. If the predicted task start time has not arrived, the program is suspended and waits for the next wake-up;
[0023] S5, start the training task;
[0024] S6, start the prediction task;
[0025] S7: Suspend the program and wait for the timer to wake up.
[0026] In a preferred solution, starting a training task includes the following steps:
[0027] S51, reading the patient's hit index and information of high-risk patients from the database;
[0028] S52, converting the read data into a table format;
[0029] S53, extract all high-risk patient information, and randomly extract non-high-risk patient information in proportion to control the proportion of no-hit patients among non-high-risk personnel;
[0030] S54, randomly shuffle the extracted data and divide them into a training set and a test set in a ratio of 75:25;
[0031] S55. Use the parameter grid model_selection.GridSearchCV of the sklearn library to select the optimal model hyperparameters, and perform training and verification;
[0032] S56. Based on the cost-sensitive error rate, the impact of missed alarms is much greater than that of false alarms. The cost matrix is adjusted to reflect the impact of missed alarms.
[0033] S57. According to the cost-sensitive error rate, feature selection is performed through the sklearn library feature_selection.SelectKBest function to obtain the optimal model;
[0034] S58, storing the parameters of the optimal model on a storage medium;
[0035] In a preferred solution, starting a prediction task includes the following steps:
[0036] S61. Use the load function of the joblib library to load model parameters;
[0037] S62, reading patient information and related indicators, and assigning low risk to patients without hit indicators;
[0038] S63, converting the patient data with hit indicators into a table format;
[0039] S64, using the loaded model to perform prediction calculation on the converted data, and calculating the probability value of high risk;
[0040] S65. Write the calculation results and low-risk patients back to the database, update the patient's risk level, and send it to the user terminal at the same time. For high-risk patients, send the results to the terminals of relevant medical staff to remind early intervention.
[0041] The beneficial effects of the present invention are:
[0042] (1) By real-time monitoring and predicting the medical risks that patients may face, hospitals can identify high-risk patients in advance and take preventive measures.
[0043] (2) The predictive model helps reduce patient complaints and disputes caused by medical errors or dissatisfaction, and improves the hospital's service quality and patient satisfaction.
[0044] (3) By promptly identifying and intervening in high-risk situations, the risk of patients suffering from serious adverse medical events can be reduced.
[0045] (4) By demonstrating that the hospital attaches great importance to patient safety and takes positive measures, patients' trust in the hospital can be enhanced.
[0046] (5) The hybrid convolutional model provides beneficial effects in feature extraction, prediction accuracy, generalization ability, etc. by combining the advantages of different models, making it a powerful tool for processing complex data sets. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 It is a structural diagram of the present invention.
[0048] Figure 2 It is a flow chart of the training module of the present invention.
[0049] Figure 3 It is a structural diagram of the prediction module in the present invention.
[0050] Figure 4 It is a diagram of the overall working steps of an embodiment of the present invention. DETAILED DESCRIPTION
[0051] Example 1
[0052] like Figure 1 As shown, the present invention is composed of a control module, a prediction module, a training module and a storage module. The control module is used to control the overall workflow. The training module obtains patient indicators, complaints, and adverse events from the database through the indicator extraction system, and performs machine learning model training. The training results are stored in the storage module. When a prediction operation is required, the prediction module obtains the model parameters from the storage module for loading, and also obtains the various indicators of the current patient from the database through the indicator extraction system, and then performs the prediction operation. After the prediction is completed, the present invention will send the prediction results to the database and the terminal through the network.
[0053] 1. Control module
[0054] The control module is mainly used for overall control of the system, which mainly includes two parts: prediction and training. When predicting, considering that the patient's indicators change slowly, this module is designed to start the prediction module every four hours through a timing strategy to ensure the update of the patient's condition. At the same time, the control module must select the model and choose the latest model for prediction. After the prediction is completed, the control module will write the prediction results back to the database and send them to the terminal device at the same time. When a high-risk patient is added, it will actively remind through text messages and other means.
[0055] For training tasks, the timing function of the control module starts a training task once a month. After the index extraction system reads the data from the database, it conducts training. After the training, the model score is judged. When the score meets the conditions, the model parameters are written into the storage module. If the score does not meet the conditions, an alarm is sent to the terminal through the network and relevant information is recorded.
[0056] 2. Training Module
[0057] The training module process is as follows Figure 2 shown.
[0058] The module first reads the patient's ID and hit indicators and other information from the database, as well as the information of high-risk patients. Then the data format is converted and stored in a table format, with one row for each ID, and each row lists the hit indicators, whether it is high-risk, and other information. When reading the database, considering that some complaints may be delayed, it is necessary to pay attention to the time of patient information to include the hospitalization time of high-risk patients. The data is stored in the format of Table 1.
[0059] Table 1 Input model data format
[0060] ID Indicator 1 Indicator 2 Indicator 3 Index High-risk patients id_1 0 1 1 …… 1 id_2 1 0 0 …… 0 id_n 0 0 1 …… 0
[0061] Currently, the proportion of high-risk patients is very low, so the data will be unbalanced. After obtaining the data information, we need to sample and extract the data, extract all the high-risk patient information, and then randomly extract the non-high-risk patient information in proportion. During the extraction, monitor the number of missed patients and control their proportion. Finally, all the data are mixed to form training data.
[0062] After obtaining the training data, the training data is grouped and randomly divided into training and test sets in a ratio of 75:25. Then the parameter grid model_selection.GridSearchCV of the sklearn library is used to select the current optimal model hyperparameters. Feature selection is performed through the sklearn library feature_selection.SelectKBest function. Scikit-learn (sklearn for short) is an open source machine learning library and one of the most widely used machine learning libraries in Python. It is built on scientific computing libraries such as NumPy, SciPy, and matplotlib, and provides a unified interface to use various machine learning algorithms. GridSearchCV is a module in the scikit-learn (sklearn) library, which is used for grid search (Grid Search) of model parameters and is an automated model tuning method. GridSearchCV can traverse a given parameter grid and find the best parameter combination through cross-validation to improve the performance of the model. SelectKBest is one of the tools for feature selection in the scikit-learn (sklearn) library. It selects the top K best features in a data set based on a given scoring function. This tool can be used to reduce the dimensionality of data and improve the performance and interpretability of models.
[0063] The machine learning model selected by the present invention is the XGBOOST model. XGBOOST is an ensemble learning algorithm that builds a powerful ensemble model by combining multiple weak learners. It is an extensible end-to-end tree augmentation system. Assuming that the data set contains n samples, each sample has m features, a model containing k trees can be expressed as:
[0064]
[0065] where F = {f(x) = ω q(x)}(q:R m ->T,ω∈R T ) is the space of regression trees, q represents the leaf subscript of each tree, T is the number of leaves, ω is the leaf weight, and each f k Corresponding to an independent tree structure.
[0066] The objective function of XGBoost is:
[0067]
[0068] in is the loss function, is the predicted value of the i-th feature at the t-th iteration, f tis the residual term, is the regularization term. Taylor expansion is performed on the objective function, the second-order derivative term is retained, and the constant term is removed. Expanding the regularization term, we get:
[0069]
[0070] where g i and h i Corresponding to the first-order derivative and the second-order derivative respectively. Decomposing the samples into samples on each node, the objective function is simplified to the sum of the unified leaf nodes, and we can get:
[0071]
[0072] Then, by taking the derivative of the objective function with respect to the leaf weight ω and substituting it into the objective function, we can get the optimal solution:
[0073]
[0074] Considering that it is impossible to enumerate all tree structures, a greedy algorithm is used in the algorithm, starting from a single leaf and iteratively adding branches to the tree. Assume that I L and I R is the instance set of the left and right nodes after splitting. After splitting:
[0075]
[0076] For each hyperparameter combination, the model results need to be evaluated through the test set. Considering the actual situation of risk warning, the impact of missed reports is far greater than that of false alarms. Therefore, the cost-sensitive error rate is used for judgment. The confusion matrix of the classification results is shown in Table 2:
[0077] Table 2 Classification results confusion matrix
[0078]
[0079] The cost matrix is shown in Table 3:
[0080] Table 3 Cost matrix
[0081]
[0082] The calculation formula of cost-sensitive error rate is:
[0083]
[0084] According to this formula, adjust the cost 01 and cost 10 , the evaluation criteria can be calculated. Since the impact of missed reports in the present invention is much greater than that of false alarms, it is necessary to adjust the cost 01>cost 10
[0085] Based on the evaluation results, the final model can be obtained and written to the disk for the prediction module to call during prediction.
[0086] 3. Prediction Module
[0087] The prediction module process is as follows Figure 3 shown.
[0088] The prediction module loads model parameters through the load function of the joblib library. Joblib is a Python open source library for efficient parallel computing. It provides easy-to-use memory mapping and parallel computing tools to distribute tasks to multiple worker processes. The joblib.load() function can load a previously saved trained model. joblib uses memory mapping file technology to efficiently load large files without loading the entire file into memory, so it is very fast and is particularly suitable for tasks that require repeated calculations or large-scale data processing.
[0089] After the data preprocessing module reads the patient information and related indicators, it preprocesses the read data and matches the hit indicators, risk level, processing time and other data according to the patient ID number. Then, the patients without hit indicators are taken out and their risk level is assigned a low risk. This preprocessing can effectively reduce the calculation time of the early warning model.
[0090] The format conversion module traverses the preprocessed data, fills in the patients and their hit indicators according to the format of Table 4, sets the hit indicators of the patients to 1, and sets the hit indicators to 0. The header indicator items need to be consistent with the format input when training the model.
[0091] Table 4 Input model data format
[0092] ID Indicator 1 Indicator 2 Indicator 3 Index id_1 0 1 1 …… id_2 1 0 0 …… id_n 0 0 1 ……
[0093] After the model reads the input data, it will perform a prediction calculation. XGBOOST performs a weighted summation on the output results of multiple base learners to obtain a more accurate prediction result. The formatted data is read into the loaded model, and the prediction is performed through the predict_proba function of the sklearn library to obtain the probability value of the final result. Based on this probability value, the patient's risk level is calculated according to the specific requirements of the hospital.
[0094] The calculation results will be published through the network and written back to the database to update the patient's risk warning results in the database. They will also be sent to the user terminal. For patients with high-risk results, the results will be sent to relevant medical staff to remind them to intervene early.
[0095] 4. Storage Module
[0096] The storage module is used to store model parameters, logs and other information generated by the system during the working process. Through Python's joblib library and logging library, the data is written to storage media such as disks.
[0097] like Figure 4 Shown is the overall flow chart of this embodiment.
[0098] Example 2
[0099] A medical safety risk assessment system includes a control module, a training module, a prediction module and a storage module;
[0100] The control module is used to control the overall workflow, including the scheduled start-up of the prediction module and the training module;
[0101] The training module is used to obtain the patient's medical indicators, complaints, and adverse events from the database, conduct model training, and store the training results;
[0102] The prediction module is used to obtain the model parameters from the storage module for loading, perform prediction operations, and send the prediction results to the database and the terminal through the network;
[0103] The storage module is used to store model parameters and log information and write data to the storage medium.
[0104] Preferably, the training module trains a hybrid neural network model of STAR-BiLSTM and XGBoost, and judges it by cost-sensitive error rate.
[0105] Preferably, in the hybrid neural network model, several intermediate layers are designed between STAR-BiLSTM and XGBoost to further process the output of STAR-BiLSTM and then use it as the input of the XGBoost model.
[0106] Preferably, the middle layer includes a summary statistics layer and an embedding layer;
[0107] Calculate statistics for the sliding window of time series data, including mean, median, maximum and minimum values, and input them into the XGBoost model as additional features;
[0108] The embedding layer is used to convert the sequence data output by STAR-BiLSTM into a dense vector to better capture the patterns in the sequence.
[0109] Preferably, the training module also preprocesses the training data, and the preprocessing process specifically includes:
[0110] Time series data: For time series data, it is necessary to ensure that the timestamps of the data are aligned and handle missing values or outliers;
[0111] Feature engineering: extracting useful features from raw data, including time series indicators and static attributes;
[0112] Data splitting: Split the training data into training set, validation set and test set.
[0113] A method for a medical safety risk assessment system comprises the following steps:
[0114] S1. Initialize the timer, design the training cycle and prediction cycle according to the hospital's requirements, and set the timer's training task and prediction task start time;
[0115] S2, start the timer;
[0116] S3, the system timer queries whether the training task start time has arrived. If the training task start time has arrived, step 5 is executed; if the training task start time has not arrived, step S5 is executed, and then query whether the predicted task start time has arrived. If the predicted task start time has arrived, step S6 is executed;
[0117] S4. If the predicted task start time has not arrived, the program is suspended and waits for the next wake-up;
[0118] S5, start the training task;
[0119] S6, start the prediction task;
[0120] S7: Suspend the program and wait for the timer to wake up.
[0121] Preferably, starting the training task includes the following steps:
[0122] S51, reading the patient's hit index and information of high-risk patients from the database;
[0123] S52, converting the read data into a table format;
[0124] S53, extract all high-risk patient information, and randomly extract non-high-risk patient information in proportion to control the proportion of no-hit patients among non-high-risk personnel;
[0125] S54, randomly shuffle the extracted data and divide them into a training set and a test set in a ratio of 75:25;
[0126] S55. Use the parameter grid model_selection.GridSearchCV of the sklearn library to select the optimal model hyperparameters, and perform training and verification;
[0127] S56. Based on the cost-sensitive error rate, the impact of missed alarms is much greater than that of false alarms. The cost matrix is adjusted to reflect the impact of missed alarms.
[0128] S57. According to the cost-sensitive error rate, feature selection is performed through the sklearn library feature_selection.SelectKBest function to obtain the optimal model;
[0129] S58, storing the parameters of the optimal model on a storage medium;
[0130] Preferably, starting the prediction task includes the following steps:
[0131] S61. Use the load function of the joblib library to load model parameters;
[0132] S62, reading patient information and related indicators, and assigning low risk to patients without hit indicators;
[0133] S63, converting the patient data with hit indicators into a table format;
[0134] S64, using the loaded model to perform prediction calculation on the converted data, and calculating the probability value of high risk;
[0135] S65. Write the calculation results and low-risk patients back to the database, update the patient's risk level, and send it to the user terminal at the same time. For high-risk patients, send the results to the terminals of relevant medical staff to remind early intervention.
[0136] Example 3
[0137] A medical safety risk assessment system to monitor and predict in real time the medical risks that patients may face, including medical complaints and adverse events. The system aims to help hospital managers and medical staff take preventive measures to reduce medical disputes and improve patient satisfaction by identifying high-risk patients in advance. The implementation steps of the system include:
[0138] 1. Data collection and integration:
[0139] Hospitals collect patient medical indicator data from electronic health record (EHR) systems, including vital signs, laboratory test results, medication use records, and historical complaint records.
[0140] The data is aggregated and stored in a central database for use in model training and prediction.
[0141] 2. System deployment and configuration:
[0142] Deploy a hybrid neural network model, including STAR-BiLSTM and XGBoost models, on the hospital’s server.
[0143] Configure the control module, set up scheduled tasks, perform model training every morning, and perform risk prediction every four hours.
[0144] 3. Model training:
[0145] The training module reads the patient data of the past 24 hours from the database, including medical indicators and historical complaints.
[0146] Use the STAR-BiLSTM layer to process time series data and extract features.
[0147] The output of STAR-BiLSTM is further processed through intermediate layers, including statistics calculation and feature conversion.
[0148] The processed features are input into the XGBoost model for model training.
[0149] During training, the cost-sensitive error rate is used to evaluate and optimize the model.
[0150] 4. Risk prediction and alert:
[0151] The prediction module loads the latest model parameters from the storage module every four hours and uses the current patient data to perform risk assessment.
[0152] The evaluation results are sent to the database and the medical staff's terminal devices via the network.
[0153] For patients assessed as high risk, the system automatically sends alerts to relevant doctors and nurses so that they can intervene in time.
[0154] 5. Result feedback and model iteration:
[0155] Medical staff intervene on patients based on the predicted results and record the intervention measures and results.
[0156] These feedback data are used for further training and optimization of the model.
[0157] The system regularly evaluates forecast accuracy and adjusts model parameters based on actual application results.
[0158] Implementation effect:
[0159] Risk warning: Hospitals can identify high-risk patients who may suffer medical complaints or adverse events in advance.
[0160] Intervention measures: Based on system alerts, medical staff took preventive intervention measures for high-risk patients, reducing the occurrence of medical disputes.
[0161] Patient Satisfaction: Through timely intervention and care, patient satisfaction with medical services is improved.
[0162] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A medical safety risk assessment system, characterized by: It includes a control module, a training module, a prediction module and a storage module; The control module is used to control the overall workflow, including the scheduled start-up of the prediction module and the training module; The training module is used to obtain the patient's medical indicators, complaints, and adverse events from the database, conduct model training, and store the training results; The prediction module is used to obtain the model parameters from the storage module for loading, perform prediction operations, and send the prediction results to the database and the terminal through the network; The storage module is used to store model parameters and log information and write data to the storage medium.
2. A medical safety risk assessment system according to claim 1, characterized in that: Build a hybrid neural network model with STAR-BiLSTM as the underlying network to extract time series features; The training module trains the hybrid neural network model of STAR-BiLSTM and XGBoost and judges it by the cost-sensitive error rate.
3. A medical safety risk assessment system according to claim 2, characterized in that: In the hybrid neural network model, several intermediate layers are designed between STAR-BiLSTM and XGBoost to further process the output of STAR-BiLSTM and then use it as the input of the XGBoost model.
4. A medical safety risk assessment system according to claim 3, characterized in that: The middle layers include summary statistics layer and embedding layer; Calculate statistics for the sliding window of time series data, including mean, median, maximum and minimum values, and input them into the XGBoost model as additional features; The embedding layer is used to convert the sequence data output by STAR-BiLSTM into a dense vector to better capture the patterns in the sequence.
5. A medical safety risk assessment system according to claim 2, characterized in that: The training module also preprocesses the training data. The preprocessing process specifically includes: Time series data: For time series data, it is necessary to ensure that the timestamps of the data are aligned and handle missing values or outliers; Feature engineering: extracting useful features from raw data, including time series indicators and static attributes; Data splitting: Split the training data into training set, validation set and test set.
6. A method for using a medical safety risk assessment system according to any one of claims 1 to 5, characterized in that: The following steps are involved: S1. Initialize the timer, design the training cycle and prediction cycle according to the hospital's requirements, and set the timer's training task and prediction task start time; S2, start the timer; S3, the system timer queries whether the training task start time has arrived. If the training task start time has arrived, step 5 is executed; if the training task start time has not arrived, step S5 is executed, and then query whether the predicted task start time has arrived. If the predicted task start time has arrived, step S6 is executed; S4. If the predicted task start time has not arrived, the program is suspended and waits for the next wake-up; S5, start the training task; S6, start the prediction task; S7: Suspend the program and wait for the timer to wake up.
7. The method of a medical safety risk assessment system according to claim 6, characterized in that: Start the training task, including the following steps: S51, reading the patient's hit index and information of high-risk patients from the database; S52, converting the read data into a table format; S53, extract all high-risk patient information, and randomly extract non-high-risk patient information in proportion to control the proportion of no-hit patients among non-high-risk personnel; S54, randomly shuffle the extracted data and divide them into a training set and a test set in a ratio of 75:25; S55. Use the parameter grid model_selection.GridSearchCV of the sklearn library to select the optimal model hyperparameters, and perform training and verification; S56. Based on the cost-sensitive error rate, the impact of missed alarms is much greater than that of false alarms. The cost matrix is adjusted to reflect the impact of missed alarms. S57. According to the cost-sensitive error rate, feature selection is performed through the sklearn library feature_selection.SelectKBest function to obtain the optimal model; S58. Store the parameters of the optimal model in a storage medium.
8. The method of a medical safety risk assessment system according to claim 6, characterized in that: Starting a prediction task includes the following steps: S61. Use the load function of the joblib library to load model parameters; S62, reading patient information and related indicators, and assigning low risk to patients without hit indicators; S63, converting the patient data with hit indicators into a table format; S64, using the loaded model to perform prediction calculation on the converted data, and calculating the probability value of high risk; S65. Write the calculation results and low-risk patients back to the database, update the patient's risk level, and send it to the user terminal at the same time. For high-risk patients, send the results to the terminals of relevant medical staff to remind early intervention.