Training method and system for predicting model of hyperamylasemia after ERCP

The method enhances hyperamylasemia prediction by data preprocessing, clustering, and label correction, improving accuracy and reliability of ERCP postoperative risk assessment using a computer system.

CN118507070BActive Publication Date: 2024-12-06AFFILIATED CHILDRENS HOSPITAL OF CAPITAL INST OF PEDIATRICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410714077.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-04
Publication Date
2024-12-06
Estimated Expiration
2044-06-04

AI Technical Summary

Technical Problem

Existing methods for predicting hyperamylasemia after ERCP rely heavily on clinical experience and single monitoring indicators, leading to subjective and low accuracy in risk assessment due to the complexity and diversity of medical data.

Method used

A method and system for training a prediction model involving data preprocessing, clustering, and label correction to enhance the reliability and accuracy of hyperamylasemia prediction, utilizing a computer system with a memory and processor to execute the training process.

Benefits of technology

The method improves the prediction accuracy of hyperamylasemia by grouping monitoring data samples, correcting labels, and training a model that can identify high-risk cases more accurately, providing valuable clinical decision support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118507070B_ABST
    Figure CN118507070B_ABST
Patent Text Reader

Abstract

The present application provides a method and system for training a prediction model for hyperamylasemia after ERCP. After determining the monitoring indicator characterization vector corresponding to a post-ERCP monitoring data sample library, the monitoring indicator characterization vector corresponding to each post-ERCP monitoring data sample in the post-ERCP monitoring data sample library is grouped to obtain a plurality of groups; and a target group with danger is selected therein to change the label of the hyperamylasemia non-occurring group sample corresponding to the monitoring indicator characterization vector in the target group, thereby completing the adjustment of the hyperamylasemia label in the post-ERCP monitoring data sample library and obtaining a revised post-ERCP monitoring data sample library, so that the samples of the training model have high reliability. Subsequently, an initial post-ERCP hyperamylasemia prediction model is trained based on the revised post-ERCP monitoring data sample library, so that the prediction accuracy of the obtained target post-ERCP hyperamylasemia prediction model can be higher.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for training a prediction model for hyperamylasemia after ERCP. Background Art

[0002] In the medical field, with the continuous development of medical technology and the application of data science, machine learning models have become a trend in assisting doctors in disease prediction and risk assessment. For example, in the postoperative management of endoscopic retrograde cholangiopancreatography (ERCP), the prediction of complications such as hyperamylasemia is of great significance for optimizing patient treatment and preventing adverse events. However, the existing prediction methods for hyperamylasemia after ERCP mainly rely on the clinical experience of doctors and the interpretation of a single monitoring indicator. This method often has the problems of strong subjectivity and low prediction accuracy. In addition, due to the complexity and diversity of medical data, how to effectively extract valuable information from a large amount of monitoring data and accurately predict the risk of hyperamylasemia in patients is an important challenge. Summary of the invention

[0003] In view of this, the embodiment of the present application at least provides a method and system for training a prediction model for hyperamylasemia after ERCP. The technical solution of the present application is implemented as follows:

[0004] On the one hand, the present application provides a method for training a prediction model for post-ERCP hyperamylasemia, the method comprising: obtaining an ERCP post-operative monitoring data sample library corresponding to a training case library; the ERCP post-operative monitoring data sample library comprises samples of a non-hyperamylasemia group and samples of a hyperamylasemia group; the samples of the non-hyperamylasemia group carry a non-hyperamylasemia marker, and the samples of the hyperamylasemia group carry a hyperamylasemia marker; determining a monitoring indicator characterization vector corresponding to each post-ERCP post-operative monitoring data sample in the ERCP post-operative monitoring data sample library; grouping according to a plurality of the monitoring indicator characterization vectors corresponding to the ERCP post-operative monitoring data sample library to obtain a plurality of groups; determining a target group based on the monitoring indicator characterization vectors respectively corresponding to the plurality of groups; correcting the markers of the non-hyperamylasemia group samples corresponding to the target group to the hyperamylasemia marker to obtain a corrected ERCP post-operative monitoring data sample library; training an initial post-ERCP hyperamylasemia prediction model based on the corrected post-ERCP monitoring data sample library to obtain a target post-ERCP hyperamylasemia prediction model.

[0005] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0006] The beneficial effects of the present application include at least: after determining the monitoring indicator characterization vector corresponding to the ERCP postoperative monitoring data sample library, the present application groups the monitoring indicator characterization vectors corresponding to each ERCP postoperative monitoring data sample in the ERCP postoperative monitoring data sample library to obtain multiple groups; and selects a target group with danger therein to change the label of the hyperamylasemia non-occurring group sample corresponding to the monitoring indicator characterization vector in the target group, completes the adjustment of the hyperamylasemia label in the ERCP postoperative monitoring data sample library, and obtains a revised ERCP postoperative monitoring data sample library, so that the samples of the training model have high reliability, and then trains the initial ERCP postoperative hyperamylasemia prediction model based on the revised ERCP postoperative monitoring data sample library, which can make the obtained target ERCP postoperative hyperamylasemia prediction model have higher prediction accuracy, and assist doctors in accurately and quickly identifying high-risk cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of the implementation flow of a method for training a prediction model for hyperamylasemia after ERCP provided in an embodiment of the present application.

[0008] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0009] The embodiment of the present application provides a method for training a prediction model for hyperamylasemia after ERCP, which can be executed by a processor of a computer system. The computer system may refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, and a desktop computer. It can be understood that the technical solution of the present application is a process of training a model and does not involve specific disease diagnosis and treatment.

[0010] Figure 1 A schematic diagram of the implementation flow of a method for training a prediction model for hyperamylasemia after ERCP provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes:

[0011] Step S100: Obtain the ERCP postoperative monitoring data sample library corresponding to the training case library. The ERCP postoperative monitoring data sample library includes samples of the hyperamylasemia non-occurrence group and samples of the hyperamylasemia occurrence group; the samples of the hyperamylasemia non-occurrence group carry the hyperamylasemia non-occurrence mark, and the samples of the hyperamylasemia occurrence group carry the hyperamylasemia occurrence mark.

[0012] In step S100, the training case library is a collection containing a large amount of historical case data, which is used to train the machine learning model so that it can learn to identify patterns or make predictions from input data. For example, in an embodiment of the present application, the training case library may contain all patient data that have undergone ERCP surgery in the past few years, including surgical records, postoperative monitoring data, diagnostic results, etc. The ERCP postoperative monitoring data sample library is a subset extracted from the training case library, specifically used for the training of the ERCP postoperative hyperamylasemia prediction model. This sample library contains monitoring data related to ERCP surgery, such as serum amylase levels, abdominal pain, surgical success rate, etc. For example, assuming that there are 10,000 ERCP surgery cases in the training case library, then the ERCP postoperative monitoring data sample library may be 5,000 cases selected from them, which contain detailed postoperative monitoring data, such as the change curve of serum amylase levels, postoperative pain scores, etc. In the ERCP postoperative monitoring data sample library, the hyperamylasemia non-occurrence group samples refer to those cases in which the serum amylase levels remain within the normal range after surgery and hyperamylasemia does not occur. These samples are usually regarded as positive samples because they conform to a certain "normal" or "no event occurred" state. For example, in the embodiment of the present application, assuming that there are 2500 cases with normal serum amylase levels after surgery and no hyperamylasemia occurs, then these 2500 cases constitute the hyperamylasemia non-occurrence group samples. In contrast to the hyperamylasemia non-occurrence group samples, the hyperamylasemia occurrence group samples refer to those cases in which the serum amylase levels are abnormally elevated after surgery and hyperamylasemia occurs. In the context of machine learning, these samples are usually regarded as negative samples because they represent a state of "abnormal" or "occurrence".

[0013] The hyperamylasemia non-occurrence marker and the hyperamylasemia occurrence marker are labels or identifiers associated with each sample in the ERCP post-operative monitoring data sample library, which are used to indicate whether the sample has hyperamylasemia. These labels are important supervisory information in the machine learning model training process, helping the model learn how to distinguish different sample categories. For example, in the ERCP post-operative monitoring data sample library, each sample is assigned a label. For samples where hyperamylasemia does not occur, the label may be "0" or "False"; for samples where hyperamylasemia occurs, the label may be "1" or "True". These labels will be used as target variables during model training, and the model will learn how to predict the values ​​of these target variables based on input features (such as postoperative serum amylase levels, abdominal pain, etc.).

[0014] Specifically, in step S100, the computer system selects data related to post-ERCP monitoring from the existing medical case library and constructs a special post-ERCP monitoring data sample library. This sample library contains two main types of data samples: samples of the group without hyperamylasemia and samples of the group with hyperamylasemia. The samples of the group without hyperamylasemia can be understood as positive samples, which represent those cases in which the serum amylase level did not increase abnormally after ERCP surgery, that is, cases in which hyperamylasemia did not occur. These samples carry a mark that hyperamylasemia did not occur, which is usually information recorded by doctors when evaluating the patient's status after surgery, such as serum amylase levels within the normal range. On the other hand, the samples of the group with hyperamylasemia can be understood as negative samples, which represent those cases in which the serum amylase level increased abnormally after ERCP surgery, that is, cases in which hyperamylasemia occurred. These samples carry a mark of hyperamylasemia occurrence, and usually contain detailed post-operative monitoring data, such as the specific value of the serum amylase level, the time point of hyperamylasemia occurrence, and the patient's clinical manifestations.

[0015] In actual operation, the computer system can extract this data from the medical case database through automated scripts or database query statements. For example, it can search for case records containing keywords such as "post-ERCP" and "serum amylase level", and then classify these cases as positive samples or negative samples based on the serum amylase level and the doctor's evaluation results.

[0016] Step S200: Determine the monitoring index representation vector corresponding to each ERCP post-operative monitoring data sample in the ERCP post-operative monitoring data sample library.

[0017] In the embodiment of the present application, in step S200, the primary task of the computer system is to analyze each data sample in the ERCP postoperative monitoring data sample library. These samples contain a wealth of postoperative monitoring information, such as serum amylase levels, abdominal pain conditions, and surgical success rates, which are crucial for predicting the occurrence of hyperamylasemia.

[0018] In order to convert these raw data into a form that can be processed by the machine learning model, the computer system uses feature extraction to convert the raw data into feature vectors, which can capture the inherent laws and key information of the data. First, the computer system identifies monitoring indicators that are closely related to the occurrence of hyperamylasemia, such as serum amylase levels at different time points after surgery, abdominal pain scores, and whether analgesics are used. These indicators are regarded as key features of the prediction model. Next, for each sample of post-ERCP monitoring data, the computer system extracts the specific values ​​of these key features. For example, for a certain sample, the serum amylase level values ​​1 hour, 6 hours, and 24 hours after surgery, and the postoperative abdominal pain score of 5 (assuming the score range is 0-10) can be extracted. Then, these extracted feature values ​​are converted into a feature vector. A feature vector is a numerical array in which each element corresponds to the specific value of a feature. Taking the above sample as an example, its feature vector may be as follows: [serum amylase level 1 hour after surgery, serum amylase level 6 hours after surgery, serum amylase level 24 hours after surgery, postoperative abdominal pain score] = [200U / L, 350U / L, 150U / L, 5]. This process is repeated for each sample in the entire ERCP postoperative monitoring data sample library, thereby generating a data set containing the feature vectors of all samples. This data set will be used as input for subsequent machine learning model training to help the model learn how to predict the occurrence of hyperamylasemia from the feature vectors. Through the feature extraction and conversion process of step S200, the computer system converts the original ERCP postoperative monitoring data into a form that can be understood and processed by the machine learning model, laying a solid foundation for the subsequent hyperamylasemia prediction model training.

[0019] Step S300: grouping according to a plurality of monitoring indicator representation vectors corresponding to the ERCP post-operative monitoring data sample library to obtain a plurality of groups.

[0020] In the embodiment of the present application, step S300 groups multiple monitoring index characterization vectors in the ERCP postoperative monitoring data sample library, i.e., cluster analysis. Specifically, in step S300, the computer system first uses a clustering algorithm (such as a K-means clustering algorithm) to process the monitoring index characterization vectors in the ERCP postoperative monitoring data sample library. These characterization vectors are extracted from step S200, and each vector contains a series of characteristic values ​​related to ERCP postoperative monitoring, such as serum amylase levels, abdominal pain scores, etc.

[0021] The working principle of the clustering algorithm is to divide the samples in the data set into several non-overlapping subsets (i.e., clusters) through iterative calculation, so that the samples in the same cluster are as similar as possible in characteristics, while the samples between different clusters are as different as possible in characteristics. In the embodiment of the present application, this means that the algorithm attempts to cluster samples with similar postoperative monitoring characteristics together to form different clusters. Taking the K-means clustering algorithm as an example, the computer system randomly selects K samples as the initial cluster centers. Then, for each sample in the data set, calculate its distance from each cluster center and assign it to the cluster corresponding to the nearest cluster center. Then, the algorithm recalculates the centroid of each cluster (i.e., the mean of all samples in the cluster) and uses the centroid as the new cluster center. This process will be iterated continuously until the cluster center no longer changes significantly or reaches a preset number of iterations.

[0022] For example, one cluster may include samples with persistently elevated serum amylase levels after surgery, which may be at high risk for hyperamylasemia, while another cluster may include samples that have recovered well after surgery and whose serum amylase levels fluctuate within the normal range.

[0023] Through the cluster analysis of step S300, the computer system can not only better understand the data structure in the ERCP postoperative monitoring data sample library, but also provide valuable information for the subsequent hyperamylasemia prediction model training. For example, the results of the cluster analysis can be used as a basis for feature selection to help determine which features are more important for predicting hyperamylasemia. In addition, cluster analysis can also help identify outliers or noise in the data, thereby improving the accuracy and robustness of the prediction model.

[0024] Step S400: Determine a target group based on monitoring indicator characterization vectors corresponding to a plurality of groups.

[0025] In the embodiment of the present application, step S400 involves determining a target group from multiple groups. The purpose of this step is to identify those sample groups that are closely related to the occurrence of hyperamylasemia so that more attention can be paid to them in the subsequent model training process. Specifically, in step S300, the computer system has divided the multiple monitoring index characterization vectors in the ERCP postoperative monitoring data sample library into several non-overlapping groups through a clustering algorithm. Each group contains samples with similar characteristics in the monitoring indicators.

[0026] In step S400, the computer system will further analyze these groups to determine the target group. The target group refers to the group containing samples that are closely related to the occurrence of hyperamylasemia. In order to determine the target group, the computer system can adopt the following strategies:

[0027] First, the computer system can analyze the incidence of hyperamylasemia in the samples in each group. Specifically, the proportion of samples in the hyperamylasemia group in each group is calculated. Generally, groups with higher incidence of hyperamylasemia are more likely to contain samples that are closely related to the occurrence of hyperamylasemia, so these groups are more likely to be identified as target groups. For example, suppose that 5 groups are obtained through cluster analysis, of which the proportion of samples in the hyperamylasemia group in the third group reaches 60%, while the proportions of other groups are all below 40%. In this case, the computer system can identify the third group as the target group.

[0028] Secondly, the computer system can also consider the monitoring indicator characteristics of the samples in the group and analyze the monitoring indicator characterization vectors of the samples in each group to identify those characteristics that are closely related to the occurrence of hyperamylasemia. For example, if the samples in a group show an abnormally high trend in key characteristics such as postoperative serum amylase levels and abdominal pain scores, then this group is more likely to be identified as a target group.

[0029] Finally, the computer system can combine multiple strategies to determine the target grouping. For example, preliminary screening can be performed based on the incidence of hyperamylasemia, and then the monitoring indicator characteristics in the screened groups can be further analyzed to ensure that the final target grouping contains enough samples that are closely related to the occurrence of hyperamylasemia and has a clear characteristic pattern.

[0030] Through the target grouping process of step S400, the computer system can identify the sample group closely related to the occurrence of hyperamylasemia in the ERCP post-operative monitoring data sample library, thereby providing more targeted data support for subsequent hyperamylasemia prediction model training.

[0031] Step S500: Correct the markers of the samples in the group without hyperamylasemia corresponding to the target group into markers of hyperamylasemia occurrence, and obtain a corrected ERCP postoperative monitoring data sample library.

[0032] In the embodiment of the present application, step S500 corrects the labeling of samples in the target group that do not have hyperamylasemia. The purpose of this step is to simulate possible misjudgments or omissions in real situations, thereby improving the generalization ability and robustness of the prediction model.

[0033] Specifically, in step S400, the computer system has determined the target groups through analysis, and these groups contain samples that are closely related to the occurrence of hyperamylasemia. However, even in the target groups, there may be some samples that are mistakenly marked as not having hyperamylasemia. These samples may be misjudged due to doctor's misjudgment, monitoring data error or other reasons.

[0034] In step S500, the computer system corrects the labels of these samples in the target group that do not have hyperamylasemia. Specifically, it corrects the labels of these samples from non-hyperamylasemia to hyperamylasemia. This correction process is based on the monitoring index characterization vector and clustering results of the samples in the target group to ensure the rationality of the correction. Taking a specific scenario as an example, assume that the target group contains 100 samples, of which 90 samples are correctly marked as hyperamylasemia, but the remaining 10 samples are incorrectly marked as hyperamylasemia. These incorrectly marked samples may have similar monitoring index characteristics to the real hyperamylasemia samples, such as abnormally elevated serum amylase levels after surgery, high abdominal pain scores, etc. In step S500, the computer system identifies these incorrectly marked samples and corrects their labels to hyperamylasemia. The corrected data sample library will contain 90 real hyperamylasemia samples and 10 samples that were originally incorrectly marked but have now been corrected, for a total of 100 negative samples.

[0035] This correction process helps simulate possible misjudgments or omissions in real situations, making the trained prediction model closer to the actual situation. By introducing these corrected samples during the training process, the prediction model will be able to learn a wider range of feature patterns and improve the prediction accuracy and generalization ability for new samples in the future. This is of great significance for the practical application of the prediction model for hyperamylasemia after ERCP.

[0036] Step S600: training an initial post-ERCP hyperamylasemia prediction model based on the modified post-ERCP monitoring data sample library to obtain a target post-ERCP hyperamylasemia prediction model.

[0037] In the embodiment of the present application, step S600 uses the modified ERCP post-operative monitoring data sample library to train the initial prediction model, and finally obtains an optimized and verified target prediction model.

[0038] In step S600, the computer system selects a suitable machine learning algorithm or neural network structure as the initial prediction model for hyperamylasemia after ERCP. This initial model can be a simple logistic regression model or a complex deep learning network, such as a multi-layer perceptron (MLP) or a convolutional neural network (CNN). The specific model selected depends on the complexity of the data, the availability of computing resources, and the requirements for model performance. Next, the computer system inputs the revised ERCP post-operative monitoring data sample library into the initial prediction model. This data sample library contains samples of hyperamylasemia that occurred and did not occur after correction in step S500, and each sample corresponds to a feature vector and a label (i.e., whether hyperamylasemia occurred).

[0039] During the model training process, the computer system iteratively adjusts the model's parameters to minimize the difference between the predicted results and the true labels. This process usually uses a loss function to measure the prediction error and gradually reduces this error through optimization algorithms such as gradient descent.

[0040] Taking the logistic regression model as an example, it attempts to find a decision boundary that can divide the sample space into two regions, so that the samples in one region mainly belong to the hyperamylasemia group, while the samples in the other region mainly belong to the non-amylasemia group. During the training process, the logistic regression model will continuously adjust its weight parameters so that the decision boundary can better fit the sample data.

[0041] As training progresses, the performance of the model on the validation set will gradually improve. The validation set is a portion of data independently divided from the original data sample library, which is used to evaluate the performance of the model on unseen data. If the performance of the model on the validation set meets the preset standards (such as accuracy, recall and other indicators meet the requirements), the training process can be stopped and an optimized target ERCP postoperative hyperamylasemia prediction model is obtained. Finally, the computer system saves the target prediction model for subsequent use in actual clinical settings. This model can automatically predict the possibility of hyperamylasemia in patients based on the input post-ERCP monitoring data, providing valuable decision support for doctors.

[0042] Through the training and optimization process of step S600, the computer system can obtain an accurate and reliable prediction model for hyperamylasemia after ERCP, providing strong technical support for the management of patients after ERCP surgery.

[0043] As an implementation manner, step S300, grouping is performed according to a plurality of monitoring indicator characterization vectors corresponding to the ERCP postoperative monitoring data sample library to obtain a plurality of groups, which may specifically include:

[0044] Step S310: determining the commonality measurement values ​​of the monitoring indicators of the non-hyperamylasemia group samples and the hyperamylasemia group samples according to the monitoring indicator characterization vectors corresponding to the non-hyperamylasemia group samples and the hyperamylasemia group samples respectively.

[0045] In an embodiment of the present application, step S310 involves analyzing the monitoring index characterization vectors of the samples of the group without hyperamylasemia and the group with hyperamylasemia to determine the commonality measurement value between them. The purpose of this step is to quantify the similarity of the two groups of samples in the monitoring indicators and provide a basis for subsequent data grouping. When implementing step S310, the computer system accesses the ERCP post-operative monitoring data sample library and extracts the monitoring index characterization vectors corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia respectively. These characterization vectors are extracted from step S200, and they contain key feature information related to post-ERCP monitoring, such as serum amylase levels, abdominal pain scores, etc., and are represented in the form of numerical arrays.

[0046] Taking serum amylase level and abdominal pain score as an example, suppose the representation vector of a sample A in the group without hyperamylasemia is [150U / L,3], and the representation vector of a sample B in the group with hyperamylasemia is [250U / L,7]. Here, "150U / L" and "250U / L" represent the values ​​of serum amylase levels after surgery, and "3" and "7" represent the values ​​of abdominal pain scores.

[0047] Next, the computer system uses a similarity measurement method to determine the commonality measurement value of sample A and sample B on the monitoring indicator. Commonly used similarity measurement methods include Euclidean distance, cosine similarity, etc. Taking Euclidean distance as an example, the formula for calculating the Euclidean distance between two vectors is:

[0048]

[0049] in, and are two vectors, Ai and Bi are corresponding elements in the vectors. Substituting the representation vectors of sample A and sample B into the formula, we can get the Euclidean distance between them, which is one of the commonality measures between them. The smaller the commonality measure, the more similar the two samples are in terms of monitoring indicators. In addition to the Euclidean distance, computer systems can also use other similarity measures such as cosine similarity to calculate commonality measures. Cosine similarity measures the similarity of two vectors by calculating the cosine value of the angle between them. The value range is [-1,1], where 1 means complete similarity and -1 means complete opposite.

[0050] Through the calculation of step S310, the computer system can obtain the commonality measurement values ​​of the monitoring indicators between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia. These commonality measurement values ​​will be used in the subsequent data grouping steps to help the computer system identify the internal structure and pattern in the data.

[0051] Step S320: grouping the common measurement values ​​of multiple monitoring indicators corresponding to the post-ERCP monitoring data sample library to obtain multiple groups.

[0052] In the embodiment of the present application, step S320 groups the commonality metric values ​​of multiple monitoring indicators calculated in the ERCP postoperative monitoring data sample library to form multiple different groups (or clusters). The purpose of this step is to classify the samples according to their similarity in monitoring indicators, providing a basis for subsequent analysis and model training. In step S320, the computer system has a set of commonality metric values ​​of monitoring indicators calculated by step S310. These commonality metric values ​​measure the similarity or distance of different ERCP postoperative monitoring data samples in monitoring indicators. For example, assuming that the Euclidean distance is used as the commonality metric value, then for two samples A and B, if the Euclidean distance between their monitoring indicator characterization vectors is small, it means that they have a high similarity in monitoring indicators. Next, the computer system will apply a clustering algorithm (such as K-means clustering algorithm) to group these commonality metric values. K-means clustering is an unsupervised learning algorithm that attempts to divide n observations into k clusters, where each observation belongs to the cluster corresponding to the nearest mean (cluster center). In the embodiment of the present application, the observed values ​​are those commonality metrics of the monitoring indicators. Assuming that k=3 is set (i.e., it is desired to divide the data into three clusters), the K-means algorithm will randomly select three commonality metrics as the initial cluster centers. Then, for each commonality metric value (i.e., each sample), its distance from the three cluster centers is calculated, and it is assigned to the cluster corresponding to the nearest cluster center. After completing the distribution of all samples, the algorithm will recalculate the mean of all commonality metrics in each cluster and use this mean as the new cluster center. This process will be iterated continuously until the cluster center no longer changes significantly or reaches the preset number of iterations. Through the K-means clustering algorithm, the computer system will eventually obtain three clusters, each of which contains a group of ERCP postoperative monitoring data samples with similar monitoring indicators. These clusters not only reveal the inherent structure and patterns in the data, but also provide valuable information for the subsequent training of hyperamylasemia prediction models. For example, a cluster may contain a large number of samples with hyperamylasemia, which indicates that the samples in this cluster have a higher risk of hyperamylasemia and thus receive more attention during model training.

[0053] Based on this, step 400, based on the monitoring indicator characterization vectors corresponding to the multiple groups, determines the target group, including:

[0054] Step S410: Determine a target group based on the common measurement values ​​of monitoring indicators corresponding to the plurality of groups.

[0055] In the embodiment of the present application, a key sub-step of step S400 is S410, i.e., determining the target group based on the commonality metric values ​​of the monitoring indicators corresponding to a plurality of groups. The purpose of this step is to identify those sample groups that are closely related to the occurrence of hyperamylasemia, so as to give more attention in subsequent model training. In step S300, the computer system has divided the samples in the ERCP postoperative monitoring data sample library into several non-overlapping groups by a clustering algorithm (such as K-means). The samples in each group have similar characteristics in the monitoring indicators, and these characteristics are measured by the commonality metric values ​​of the monitoring indicators. The commonality metric value can be the Euclidean distance between samples, cosine similarity, etc., which reflect the similarity of the samples in the feature space.

[0056] The execution process of step S410 is as follows: the computer system calculates statistics of the commonality metric values ​​of samples in each group, such as the mean or median, to represent the overall similarity of the group. These statistics can help identify groups with higher commonality metric values ​​between samples.

[0057] Next, the computer system compares the commonality metric statistics between different groups. In this process, a threshold can be set to filter out groups with commonality metric statistics higher than the threshold. The samples in these groups have high similarity in monitoring indicators and may be closely related to the occurrence of hyperamylasemia.

[0058] For example, suppose that three groups G1, G2, and G3 are obtained through the clustering algorithm, and the commonality metric statistics of the samples in each group are 0.6, 0.8, and 0.5 respectively (here it is assumed that the smaller the commonality metric value, the higher the similarity between samples). If the threshold is set to 0.7, then groups G1 and G3 with commonality metric statistics below the threshold will be excluded, and group G2 with a commonality metric statistic of 0.8 will be determined as the target group. This means that the samples in group G2 have higher similarity in monitoring indicators and are more likely to be associated with the occurrence of hyperamylasemia.

[0059] As another implementation, step S400, determining the target group based on the monitoring indicator characterization vectors corresponding to the multiple groups, may specifically include:

[0060] Step S410a: Determine a group including a danger monitoring indicator characterization vector among multiple groups as a target group.

[0061] In step S410a, the computer system reviews the results of grouping the monitoring indicator characterization vectors in the ERCP post-operative monitoring data sample library by a clustering algorithm (such as K-means). These groups represent different subsets in the data, and the samples in each subset have similar characteristics in terms of monitoring indicators. Next, the computer system will focus on a specific analysis process that aims to identify "dangerous monitoring indicator characterization vectors" that are closely related to the occurrence of hyperamylasemia. This process includes extracting data from a preset ERCP post-operative monitoring database, determining the monitoring indicator characterization vectors for each monitoring data, and grouping them by a clustering algorithm. Then, the percentage of disease occurrence monitoring data in each group is calculated, and the group with a percentage higher than a preset threshold is defined as a "dangerous group." The monitoring indicator characterization vectors in these dangerous groups are identified as "dangerous monitoring indicator characterization vectors" that are closely related to the occurrence of hyperamylasemia.

[0062] In step S410a, the computer system will use these risk monitoring indicator characterization vectors to determine the target group. Specifically, the multiple groups previously obtained by the clustering algorithm are checked to identify the groups containing the risk monitoring indicator characterization vectors. The samples in these groups show characteristic patterns that are closely related to the occurrence of hyperamylasemia in terms of monitoring indicators, and are therefore determined as target groups.

[0063] Through the target grouping process in step S410a, the computer system can accurately identify the sample group that is closely related to the occurrence of hyperamylasemia, providing targeted data support for subsequent model training. This helps to improve the prediction accuracy and generalization ability of the model, so that the model can better assist doctors in risk assessment and management of patients after ERCP in practical applications.

[0064] Among them, the hazard monitoring index characterization vector is determined through the following process:

[0065] Step S401: Acquire a preset ERCP post-operative monitoring database corresponding to a preset case database; the preset ERCP post-operative monitoring database includes monitoring data before and after the occurrence of symptoms.

[0066] In the embodiment of the present application, step S401 involves obtaining a preset ERCP postoperative monitoring database corresponding to a preset case library. This database contains a wealth of ERCP postoperative monitoring data, which is crucial for the subsequent identification of the "risk monitoring indicator characterization vector" closely related to the occurrence of hyperamylasemia. First, the computer system needs to access a preset case library. This case library is usually a large medical database containing a large number of ERCP surgical cases and their related monitoring data. These case data are carefully sorted and organized to facilitate subsequent data analysis and model training. In the preset case library, the computer system pays special attention to data related to ERCP postoperative monitoring and extracts a preset ERCP postoperative monitoring database. This database contains various monitoring data of patients after ERCP surgery, such as serum amylase levels, abdominal pain scores, surgical success rates, etc. These data are of great significance for evaluating the patient's postoperative recovery and predicting the occurrence of hyperamylasemia. Importantly, the preset ERCP postoperative monitoring database contains not only case data in which hyperamylasemia did not occur after surgery (symptom non-occurrence monitoring data), but also case data in which hyperamylasemia occurred after surgery (symptom occurrence monitoring data). These different categories of data are crucial for identifying characteristic patterns that are strongly associated with the development of hyperamylasemia.

[0067] Through the data preparation process of step S401, the computer system obtains a preset database containing rich post-ERCP monitoring data. This database provides a basis for subsequent data analysis, enabling the computer system to identify characteristic patterns closely related to the occurrence of hyperamylasemia, thereby providing strong support for the training of the prediction model for post-ERCP hyperamylasemia.

[0068] Step S402: Determine the monitoring index characterization vector corresponding to each preset ERCP post-operative monitoring data in the preset ERCP post-operative monitoring database.

[0069] In the embodiment of the present application, step S402 involves determining the monitoring index characterization vector corresponding to each preset ERCP postoperative monitoring data in the preset ERCP postoperative monitoring database. In step S402, the computer system accesses the preset ERCP postoperative monitoring database, which contains a large amount of ERCP postoperative monitoring data, which may include multiple monitoring indicators such as serum amylase level, abdominal pain score, and operation time. Each monitoring data sample represents a postoperative monitoring record of a patient. Then, the computer system analyzes each monitoring data sample in the database and extracts key feature information related to ERCP postoperative monitoring. These feature information may include specific values ​​of various monitoring indicators, such as serum amylase levels 1 hour, 6 hours and 24 hours after surgery, changes in postoperative abdominal pain scores, etc. These feature information is of great significance for predicting the occurrence of hyperamylasemia. In order to convert these feature information into a format that can be processed by a machine learning model, the computer system uses feature extraction technology to generate a monitoring index characterization vector. Feature extraction is a process of converting raw data into feature vectors, which can capture the inherent laws and key information of the data.

[0070] In the embodiment of the present application, feature extraction may involve the following steps:

[0071] Feature selection: The computer system selects monitoring indicators that are closely related to the occurrence of hyperamylasemia as features. For example, serum amylase level, abdominal pain score, operation time, etc. may be important features.

[0072] Feature encoding: For continuous features (such as serum amylase levels), the computer system may directly use its value as an element in the feature vector. For categorical features (such as abdominal pain score), the computer system may use one-hot encoding or label encoding to convert it into a numerical vector.

[0073] Feature scaling: In order to eliminate the impact of dimensions and value ranges between different features, the computer system may use methods such as standardization or normalization to scale the feature vector.

[0074] For example, suppose that a monitoring data sample in the preset ERCP postoperative monitoring database contains information that the serum amylase level is 200U / L 1 hour after surgery, the postoperative abdominal pain score is 4 (assuming the score range is 1-10), and the operation time is 60 minutes. After feature extraction, the monitoring indicator representation vector of this sample may be [200, 0, 0, 1, 0, 0, 0, 0, 4, 60], where the first 10 elements represent the unique hot encoding of the abdominal pain score (assuming that the score is 4 and the corresponding 4th position is 1, and the rest are 0), the 11th element represents the serum amylase level, and the 12th element represents the operation time (appropriately scaled). Through the processing of step S402, the computer system converts each monitoring data sample in the preset ERCP postoperative monitoring database into a monitoring indicator representation vector, which will be used for subsequent analysis and model training processes.

[0075] Step S403: grouping according to a plurality of preset monitoring index characterization vectors corresponding to a preset post-ERCP monitoring database to obtain a plurality of preset groups.

[0076] In the embodiment of the present application, step S403 involves grouping multiple preset monitoring indicator representation vectors in the preset ERCP postoperative monitoring database to form multiple preset groups. In step S403, the computer system accesses the preset ERCP postoperative monitoring database and obtains multiple preset monitoring indicator representation vectors therefrom. These vectors are extracted from the original monitoring data by step S402, and each vector contains a series of characteristic values ​​related to ERCP postoperative monitoring, such as serum amylase level, abdominal pain score, etc.

[0077] Next, the computer system applies a clustering algorithm (such as K-means clustering algorithm) to group these preset monitoring indicator representation vectors. Clustering algorithm is an unsupervised learning method that attempts to divide the samples in the data set into several non-overlapping subsets (i.e. clusters) so that the samples in the same cluster are as similar as possible in characteristics, while the samples in different clusters are as different as possible in characteristics.

[0078] In the present application embodiment, the goal of the clustering algorithm is to cluster samples with similar monitoring index characteristics to form different preset groups. These groups may reflect different postoperative recovery states or risk levels, which is helpful for the identification of hyperamylasemia risk in subsequent analysis.

[0079] Taking the K-means clustering algorithm as an example, the computer system randomly selects K preset monitoring indicator representation vectors as the initial cluster centers. Then, for each preset monitoring indicator representation vector in the data set, calculate its distance from each cluster center (such as Euclidean distance), and assign it to the cluster corresponding to the nearest cluster center. After completing the assignment of all samples, the algorithm will recalculate the mean of all samples in each cluster and use this mean as the new cluster center. This process will be iterated continuously until the cluster center no longer changes significantly or reaches the preset number of iterations. Through the K-means clustering algorithm or other similar clustering algorithms, the computer system will eventually obtain multiple preset groups, each of which contains a group of preset ERCP postoperative monitoring data samples with similar monitoring indicators. These preset groups provide the basis for identifying dangerous monitoring indicator representation vectors in subsequent steps, which helps the computer system to more accurately identify sample characteristics closely related to the occurrence of hyperamylasemia.

[0080] For example, suppose there are 100 samples in the preset ERCP postoperative monitoring database, and three preset groups are obtained through the K-means clustering algorithm. Among them, the first group contains samples with persistently elevated serum amylase levels and high abdominal pain scores after surgery, which may reflect a higher risk of hyperamylasemia; the second group contains samples with good postoperative recovery, with serum amylase levels and abdominal pain scores within the normal range; the third group contains samples with other characteristics. These preset groups provide important reference information for determining the characterization vector of risk monitoring indicators in subsequent steps.

[0081] Step S404: Determine the percentage of symptom occurrence monitoring data corresponding to each preset group.

[0082] In the embodiment of the present application, step S404 is an important data analysis step, which involves determining the percentage of the disease occurrence monitoring data (i.e., the case data of postoperative hyperamylasemia) in each preset group. The purpose of this step is to quantitatively evaluate the correlation between each group and the occurrence of hyperamylasemia, so as to help identify the "risk group" closely related to the risk of hyperamylasemia.

[0083] In step S404, the computer system reviews the multiple preset groups previously obtained by clustering algorithms (such as K-means). Each preset group contains a group of ERCP postoperative monitoring data samples that are similar in monitoring indicators. These samples may represent different postoperative recovery states or risk levels. Next, the computer system analyzes the monitoring data samples in each preset group, especially focusing on those samples marked as symptom occurrence (i.e., postoperative hyperamylasemia). For each preset group, the computer system counts the number of symptom occurrence monitoring data and calculates the percentage of this number to the total number of samples in the group. This percentage reflects the correlation between the samples in the group and the occurrence of hyperamylasemia.

[0084] For example, suppose that three preset groups G1, G2, and G3 are obtained through the clustering algorithm. In the G1 group, there are 100 samples, of which 10 samples are symptom occurrence monitoring data; in the G2 group, there are 80 samples, of which 30 samples are symptom occurrence monitoring data; in the G3 group, there are 120 samples, of which 15 samples are symptom occurrence monitoring data. Then, the percentages of symptom occurrence monitoring data corresponding to the G1, G2, and G3 groups are 10%, 37.5%, and 12.5%, respectively.

[0085] By calculating the percentage of disease occurrence monitoring data corresponding to each preset grouping, the computer system is able to quantitatively evaluate the correlation between each grouping and the occurrence of hyperamylasemia. The higher the percentage, the more likely the samples in the group are to be associated with the occurrence of hyperamylasemia. This provides an important basis for identifying "risk groups" in subsequent steps. In the example above, the G2 grouping has the highest percentage (37.5%), so it is more likely to be identified as a "risk group" that is closely related to the risk of hyperamylasemia.

[0086] Step S405: Determine the preset group whose percentage of symptom occurrence monitoring data is greater than the percentage threshold as a risk group.

[0087] In the embodiment of the present application, step S405 is a key screening step, which involves determining the preset groups whose percentage of symptom occurrence monitoring data is greater than the preset percentage threshold as "risk groups". The purpose of this step is to identify those groups that are closely related to the occurrence of hyperamylasemia, so that these groups can be analyzed and processed in more detail in subsequent steps.

[0088] In step S405, the computer system reviews the percentage of disease occurrence monitoring data corresponding to each preset group calculated in the previous step. This percentage reflects the correlation between the samples in the group and the occurrence of hyperamylasemia. The higher the percentage, the more likely the samples in the group are related to the occurrence of hyperamylasemia. Next, the computer system sets a percentage threshold. This threshold is determined according to the actual application scenario and needs, and is used to screen out groups that are highly related to the occurrence of hyperamylasemia. The setting of the threshold usually needs to consider factors such as the distribution of data, the prediction performance requirements of the model, and the actual clinical significance. Then, the computer system compares the percentage of disease occurrence monitoring data of each preset group with the set percentage threshold. If the percentage of a group is greater than the threshold, then the group is determined as a "risk group". The samples in the risk group show characteristic patterns that are closely related to the occurrence of hyperamylasemia in monitoring indicators, so they are the objects that need to be focused on and analyzed in subsequent steps.

[0089] Through the screening process of step S405, the computer system can accurately identify the preset groups that are closely related to the occurrence of hyperamylasemia, providing important data support for the determination of the subsequent risk monitoring indicator characterization vector and model training. This helps to improve the prediction accuracy and generalization ability of the model, so that the model can better assist doctors in risk assessment and management of patients after ERCP in practical applications.

[0090] Step S406: Determine the preset monitoring indicator characterization vector in the risk group as the risk monitoring indicator characterization vector.

[0091] In step S406, the computer system reviews the risk groups determined in the previous steps. These risk groups are determined by comparing the percentage of symptom occurrence monitoring data of each preset group with the preset percentage threshold, which includes samples closely related to the occurrence of hyperamylasemia. Next, the computer system extracts the preset monitoring indicator characterization vectors corresponding to each sample in these risk groups. These characterization vectors are generated in step S402 and contain key feature information related to post-ERCP monitoring, such as serum amylase levels, abdominal pain scores, etc. Since the samples in the risk groups are closely related to the occurrence of hyperamylasemia, their corresponding monitoring indicator characterization vectors are likely to contain key features related to the risk of hyperamylasemia. Then, the computer system determines these preset monitoring indicator characterization vectors extracted from the risk groups as risk monitoring indicator characterization vectors.

[0092] Through the determination process of step S406, the computer system can accurately identify the monitoring indicator characteristics that are closely related to the occurrence of hyperamylasemia, providing strong data support for subsequent target grouping determination and model training. This helps to improve the prediction accuracy and generalization ability of the model, so that the model can better assist doctors in risk assessment and management of patients after ERCP in practical applications.

[0093] As an implementation manner, step S403, grouping is performed according to a plurality of preset monitoring index characterization vectors corresponding to a preset ERCP postoperative monitoring database to obtain a plurality of preset groups, which may specifically include:

[0094] Step S4031: determining the commonality measurement value of the preset monitoring indicators of the monitoring data without symptom occurrence and the monitoring data with symptom occurrence according to the preset monitoring indicator characterization vectors respectively corresponding to the monitoring data without symptom occurrence and the monitoring data with symptom occurrence.

[0095] In the embodiment of the present application, step S4031 involves calculating the commonality metric value of the non-symptom monitoring data and the symptom-occurring monitoring data in the preset ERCP postoperative monitoring database. The purpose of this step is to quantitatively evaluate the similarity or difference of different monitoring data samples in monitoring indicators.

[0096] Specifically, in step S4031, the computer system accesses the preset ERCP post-operative monitoring database, and extracts the preset monitoring indicator characterization vectors corresponding to the monitoring data of the symptoms that have not occurred and the monitoring data of the symptoms that have occurred. These characterization vectors are generated from the original monitoring data through a specific feature extraction algorithm, and they contain key feature information related to post-ERCP monitoring, such as serum amylase levels, abdominal pain scores, etc. Then, the computer system calculates these characterization vectors to determine the commonality metric value of the preset monitoring indicators between the monitoring data of the symptoms that have not occurred and the monitoring data of the symptoms that have occurred. The commonality metric value is an indicator that measures the similarity or correlation between vectors, which can be obtained by calculating the distance (such as Euclidean distance) or similarity coefficient (such as cosine similarity) between vectors.

[0097] Taking Euclidean distance as an example, suppose there are two representation vectors A and B, representing a monitoring data of a disease not occurring and a monitoring data of a disease occurring, respectively. The elements in vectors A and B correspond to the same monitoring indicators, such as serum amylase level and abdominal pain score. By calculating the commonality metrics between all disease-free monitoring data and disease-occurring monitoring data, the computer system can quantitatively evaluate the similarity relationship between different samples. These commonality metrics will be used in the subsequent data grouping steps to help the computer system identify sample groups with similar monitoring indicator characteristics and provide targeted data support for subsequent model training.

[0098] In the embodiment of the present application, the implementation of step S4031 helps the computer system to better understand the inherent structure and pattern in the monitoring data, thereby improving the accuracy and generalization ability of the prediction model. By quantitatively evaluating the similarities or differences between different monitoring data samples, the computer system can more accurately identify the characteristic patterns closely related to the occurrence of hyperamylasemia, laying a solid foundation for subsequent target grouping and model training.

[0099] Step S4032: Grouping is performed based on the common measurement values ​​of multiple preset monitoring indicators corresponding to the preset ERCP postoperative monitoring database to obtain multiple preset groups.

[0100] In the embodiment of the present application, step S4032 is based on the commonality measurement values ​​of multiple preset monitoring indicators in the preset post-ERCP monitoring database to obtain multiple preset groups. The purpose of this step is to classify the samples according to their similarity in monitoring indicators, so as to facilitate subsequent analysis and identification of "risk monitoring indicator characterization vectors" closely related to the occurrence of hyperamylasemia.

[0101] Specifically, in step S4031, the computer system has calculated the commonality metric values ​​of the preset monitoring indicators corresponding to each monitoring data sample in the preset ERCP postoperative monitoring database. These commonality metric values ​​measure the similarity or difference of different samples in monitoring indicators, providing a basis for data grouping.

[0102] In step S4032, the computer system will apply a clustering algorithm (such as K-means clustering algorithm) to group these commonality metrics. Clustering algorithm is an unsupervised learning method that attempts to divide the samples in the data set into a number of non-overlapping subsets (i.e., clusters) so that the samples in the same cluster are as similar as possible in characteristics, while the samples in different clusters are as different as possible in characteristics.

[0103] Taking the K-means clustering algorithm as an example, the computer system first randomly selects K commonality metrics as the initial cluster centers. Then, for each preset monitoring indicator commonality metric value (i.e., each sample), calculate its distance from each cluster center (such as Euclidean distance) and assign it to the cluster corresponding to the nearest cluster center. After completing the assignment of all samples, the algorithm will recalculate the mean of all commonality metrics in each cluster and use this mean as the new cluster center. This process will continue to iterate until the cluster center no longer changes significantly or the preset number of iterations is reached.

[0104] Through the K-means clustering algorithm or other similar clustering algorithms, the computer system will eventually obtain multiple preset groups, each of which contains a group of post-ERCP monitoring data samples with similar monitoring indicators. These preset groups provide a basis for identifying dangerous monitoring indicator characterization vectors in subsequent steps, which helps the computer system to more accurately identify sample features that are closely related to the occurrence of hyperamylasemia.

[0105] At this time, the method provided by the embodiment of the present application may also include: Step S4033: determining the commonality measurement value of the selected monitoring indicators in the danger group as the commonality measurement threshold; the commonality measurement value of the selected monitoring indicators in the danger group is less than the first commonality measurement value; the first commonality measurement value is the commonality measurement value in the danger group except the commonality measurement value of the selected monitoring indicators.

[0106] In the embodiment of the present application, step S4033 further determines a commonality measurement threshold based on the grouping result of step S403. This threshold will be used in subsequent steps to screen monitoring data related to the target group, helping the computer system to more accurately identify sample features closely related to the occurrence of hyperamylasemia.

[0107] In step S4032, the computer system has divided the monitoring data samples in the preset ERCP postoperative monitoring database into multiple preset groups through a clustering algorithm (such as K-means), among which some groups may contain samples that are closely related to the occurrence of hyperamylasemia, and these groups are called "risk groups".

[0108] The goal of step S4033 is to select a suitable commonality metric value from the risk group as the commonality metric threshold. This threshold will be used in subsequent steps to screen monitoring data related to the target group.

[0109] Specifically, the computer system analyzes the commonality metrics of the monitoring indicators corresponding to each sample in the risk group. These commonality metrics measure the similarity or difference of the samples in the monitoring indicators. In the risk group, the commonality metrics between samples are usually small, indicating that they have a high similarity in the monitoring indicators.

[0110] Next, the computer system selects a specific commonality metric value from the risk group as the commonality metric value of the selected monitoring indicator. This selected value needs to be smaller than other commonality metric values ​​in the risk group (i.e., the first commonality metric value). The selection of the selected value can be based on a variety of strategies, such as selecting the median, the lower quartile, or a fixed proportion of quantiles. Selecting a smaller selected value helps ensure that the target groupings screened in subsequent steps have higher similarity and relevance to the occurrence of hyperamylasemia.

[0111] Once the commonality metric value of the monitoring indicator is selected, the computer system will determine it as the commonality metric threshold. This threshold will be used in subsequent steps to filter monitoring data related to the target group. Specifically, when the computer system groups new monitoring data samples, it calculates the commonality metric values ​​of the samples in each group and compares these values ​​with the commonality metric threshold. If the commonality metric value of the samples in a group is less than or equal to the commonality metric threshold, then the group is likely to be identified as a target group, indicating that the samples in the group are closely related to the occurrence of hyperamylasemia.

[0112] For example, it is assumed that three preset groups G1, G2 and G3 are obtained through step S4032, and G2 is determined as a risk group. In G2, the commonality metric values ​​between samples range from 0.1 to 0.5 (the smaller the value, the higher the similarity). The computer system selects 0.2 as the commonality metric value of the selected monitoring indicator and determines it as the commonality metric threshold. When grouping new monitoring data samples, if the commonality metric value of the samples in a group is less than or equal to 0.2, then the group is likely to be identified as a target group, indicating that the samples in the group are closely related to the occurrence of hyperamylasemia. By determining the commonality metric threshold through step S4033, the computer system can more accurately screen out the monitoring data related to the target group, and provide targeted data support for the subsequent hyperamylasemia prediction model training. This helps to improve the prediction accuracy and generalization ability of the model, so that the model can better assist doctors in risk assessment and management of patients after ERCP in practical applications.

[0113] Step S410, determining a target group based on the commonality measurement values ​​of the monitoring indicators corresponding to the multiple groups, may specifically include:

[0114] Step S411: determine the common measurement value of the target monitoring indicator in each group; the common measurement value of the target monitoring indicator in any group is less than the second common measurement value in any group; the second common measurement value in any group is the common measurement value in any group except the common measurement value of the target monitoring indicator.

[0115] Specifically, in step S411, the computer system analyzes the commonality metrics of the monitoring indicators in each group. These commonality metrics are obtained by calculating the distance or similarity coefficient between the monitoring indicator representation vectors corresponding to the samples in the group, and they measure the similarity or difference of the samples in the monitoring indicators.

[0116] For each group, the computer system needs to determine a commonality metric value for the target monitoring indicator. This value is usually selected as the minimum value among the other commonality metric values ​​in the group except the maximum value, that is, the second commonality metric value. Selecting such a value as the commonality metric value for the target monitoring indicator helps ensure that the samples in the target group have relatively high similarity in the monitoring indicator, while excluding those samples that are too abnormal or outliers.

[0117] For example, suppose that three groups G1, G2 and G3 are obtained through the clustering algorithm. For group G1, it contains 10 samples, and the commonality measurement values ​​of the monitoring indicators corresponding to these samples are 0.1, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.6 and 0.7 respectively. In this group, the largest commonality measurement value is 0.7, and the minimum value of other commonality measurement values ​​except the maximum value is 0.2. Therefore, for group G1, its target monitoring indicator commonality measurement value is determined to be 0.2. Similarly, for groups G2 and G3, the computer system will also determine their target monitoring indicator commonality measurement values ​​respectively. These values ​​will be used to screen the monitoring data related to the target group in subsequent steps.

[0118] Through the determination process of step S411, the computer system can quantitatively evaluate the similarity of the samples in each group in terms of monitoring indicators, and identify those sample groups that are closely related to the occurrence of hyperamylasemia. This provides strong data support for subsequent target group determination and model training, and helps to improve the prediction accuracy and generalization ability of the model. In practical applications, this will help doctors better assess the risk status of patients after ERCP, thereby formulating more personalized treatment plans.

[0119] Step S412: Determine the group whose target monitoring indicator commonality metric value is not less than the commonality metric threshold as the target group.

[0120] In step S412, the computer system reviews the commonality measurement value and commonality measurement threshold of the target monitoring indicator determined in the previous step. The commonality measurement value of the target monitoring indicator is the minimum value among the other commonality measurement values ​​in each group except the maximum value, which measures the similarity of the samples in the group in terms of monitoring indicators. The commonality measurement threshold is determined based on the commonality measurement value of the selected monitoring indicator in the risk group, and is used to screen the sample group that is closely related to the occurrence of hyperamylasemia. Next, the computer system compares the commonality measurement value of the target monitoring indicator of each group with the commonality measurement threshold one by one. If the commonality measurement value of the target monitoring indicator of a group is not less than the commonality measurement threshold, then this group is determined as the target group. This means that the samples in the group have relatively high similarity in monitoring indicators and are closely related to the occurrence of hyperamylasemia.

[0121] For example, suppose that three groups G1, G2, and G3 are obtained through the previous steps, and their target monitoring indicator commonality measurement values ​​are 0.3, 0.1, and 0.4, respectively, and the commonality measurement threshold is set to 0.2. During the comparison process, it is found that the target monitoring indicator commonality measurement values ​​(0.3 and 0.4) of groups G1 and G3 are not less than the commonality measurement threshold (0.2), so they are determined as target groups. The target monitoring indicator commonality measurement value (0.1) of group G2 is less than the commonality measurement threshold, so it is not considered as a target group.

[0122] Through the screening process of step S412, the computer system can accurately identify the sample group closely related to the occurrence of hyperamylasemia, namely the target group. The samples in these groups have a high degree of similarity in monitoring indicators and are closely related to the occurrence of hyperamylasemia. This provides targeted data support for subsequent model training, which helps to improve the prediction accuracy and generalization ability of the model. In practical applications, this will help doctors better assess the risk status of patients after ERCP, so as to develop more personalized treatment plans.

[0123] In some embodiments, the method may further include:

[0124] Step S10: Determine the data item sequences corresponding to the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia.

[0125] In the embodiment of the present application, step S10 involves determining the data item sequences corresponding to the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia. This step is to extract and analyze the time series data related to postoperative monitoring in order to more comprehensively understand the recovery and risk changes of postoperative patients.

[0126] In step S10, the computer system accesses the ERCP postoperative monitoring data sample library and extracts the postoperative monitoring data of each sample therefrom. These monitoring data usually include records of multiple time points, such as serum amylase levels, abdominal pain scores and other indicators 1 hour, 6 hours, 24 hours after surgery. These records are arranged in chronological order to form a data item sequence that can reflect the dynamic changes in the patient's postoperative recovery. For samples in the group where hyperamylasemia does not occur, the computer system will extract their corresponding data item sequences. These data item sequences usually include continuous records of monitoring indicators over a period of time after surgery, reflecting the good postoperative recovery of the patient. For example, a data item sequence of a positive sample may show that the serum amylase level gradually decreases and remains within the normal range after surgery, and the abdominal pain score also gradually decreases.

[0127] Similarly, for samples of the hyperamylasemia group, the computer system will also extract their corresponding data item sequences. These data item sequences usually reflect the risk of hyperamylasemia in patients after surgery. For example, a data item sequence of a negative sample may show that the serum amylase level is abnormally elevated after surgery, and the abdominal pain score is also continuously high or shows an upward trend.

[0128] Through the processing of step S10, the computer system can obtain the data item sequences corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia. These data item sequences provide a basis for analyzing the similarities and differences between the two groups of samples in the monitoring indicators in the subsequent steps. For example, in the subsequent steps, the computer system can evaluate the similarity between the two groups of samples in the postoperative recovery status and risk changes by calculating the similarity metric between the data item sequences (such as the dynamic time warping DTW algorithm). For example, suppose there is a positive sample and a negative sample. The data item sequence of the positive sample is [serum amylase level: 150U / L, abdominal pain score: 2], which represents the monitoring data at a certain time point after surgery; while the data item sequence of the negative sample may be [serum amylase level: 300U / L, abdominal pain score: 6, serum amylase level: 350U / L, abdominal pain score: 7], which represents the continuous monitoring data at multiple time points after surgery, showing the abnormal increase trend of serum amylase level and abdominal pain score. By comparing these data item sequences, the computer system can more accurately identify the sample features that are closely related to the occurrence of hyperamylasemia.

[0129] Step S20: determining the data item commonality measurement values ​​of the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia according to the data item sequences corresponding to the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia.

[0130] In the embodiment of the present application, step S20 calculates the commonality metric value of the data items between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia, respectively. The purpose of this step is to quantitatively evaluate the similarity or difference between the two groups of samples in the time series data, and to provide an important basis for subsequent data grouping and model training. In step S20, the computer system reviews the data item sequences of the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia extracted in step S10. These data item sequences contain monitoring data at multiple time points after surgery, such as serum amylase levels, abdominal pain scores, etc., which can reflect the dynamic changes of the patient's postoperative recovery status. Then, the computer system uses a similarity measurement algorithm to calculate the commonality metric value between the two groups of sample data item sequences. This commonality metric value is used to measure the similarity or difference between the two groups of samples in the time series data. Commonly used similarity measurement algorithms include dynamic time warping (Dynamic Time Warping, DTW), LCS algorithm (Longest Common Subsequence, LCS), etc. These algorithms can handle comparisons between sequences of unequal lengths and find an optimal alignment to calculate the similarity between sequences.

[0131] Taking the dynamic time warping (DTW) algorithm as an example, the computer system performs dynamic programming calculations on the data item sequences of the two groups of samples to find an optimal alignment path so that the cumulative distance between the two groups of sequences is minimized. This cumulative distance is the data item commonality measure between the two groups of samples, which reflects the degree of similarity between the two groups of samples in the time series data. If the commonality measure is small, it means that the two groups of samples have a high similarity in the time series data; conversely, if the commonality measure is large, it means that the two groups of samples have a large difference in the time series data. Through the calculation of step S20, the computer system can obtain the data item commonality measure between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia. This commonality measure will serve as an important reference in the subsequent steps to evaluate the similarities or differences between different sample groups in the time series data, thereby helping the computer system to more accurately identify sample features that are closely related to the occurrence of hyperamylasemia.

[0132] For example, suppose there are two samples A and B, representing the sample of the group without hyperamylasemia and the sample of the group with hyperamylasemia. The data item sequence of sample A is [serum amylase level: 150U / L, abdominal pain score: 2], and the data item sequence of sample B is [serum amylase level: 300U / L, abdominal pain score: 6, serum amylase level: 350U / L, abdominal pain score: 7]. The commonality measure of the data items between them is calculated by the DTW algorithm, which is d. If the d value is small, it means that samples A and B have high similarity in time series data, which may reflect the good postoperative recovery of patients; if the d value is large, it means that samples A and B have large differences in time series data, which may represent different situations of good postoperative recovery and hyperamylasemia. This d value will be used for data grouping and model training in subsequent steps to help the computer system more accurately predict the risk of hyperamylasemia after ERCP.

[0133] At this time, step S320, grouping the common measurement values ​​of multiple monitoring indicators corresponding to the ERCP postoperative monitoring data sample library to obtain multiple groups, specifically can include: grouping the common measurement values ​​of multiple monitoring indicators and the common measurement values ​​of multiple data items corresponding to the ERCP postoperative monitoring data sample library to obtain multiple groups.

[0134] In an embodiment of the present application, step S320 groups samples in the ERCP postoperative monitoring data sample library based on multiple monitoring index commonality metrics and multiple data item commonality metrics. The purpose of this step is to group samples with similar monitoring index characteristics and time series data characteristics together to form different groups for subsequent analysis and model training. In step S320 based on a derivative implementation method, the computer system reviews the multiple monitoring index commonality metrics and multiple data item commonality metrics calculated in the previous steps. The monitoring index commonality metrics measure the similarities or differences of different samples in monitoring indicators (such as serum amylase levels, abdominal pain scores, etc.), while the data item commonality metrics measure the similarities or differences of different samples in time series data (such as monitoring data sequences at different time points after surgery).

[0135] Next, the computer system will use a clustering algorithm (such as the K-means clustering algorithm) to jointly group these commonality metrics. During the clustering process, the algorithm will consider both the commonality metrics of the monitoring indicators and the commonality metrics of the data items, and find a grouping method that makes the samples in the same group have high similarity in both dimensions. Specifically, the clustering algorithm will iteratively calculate the distance from each sample to each cluster center (based on a weighted combination of the commonality metrics of the monitoring indicators and the commonality metrics of the data items), and assign the samples to the cluster clusters corresponding to the cluster centers closest to them. As the iterations proceed, the cluster centers will gradually stabilize, forming the final multiple groups.

[0136] Through the grouping process of step S320, the computer system can group samples with similar characteristics together to form different groups. These groups provide important data support for subsequent target group determination and model training. In the target group determination step, the computer system can further screen out target groups that are closely related to the occurrence of hyperamylasemia based on the commonality metrics and clinical significance of the samples in the group. Then, based on the data of these target groups, a more accurate prediction model for hyperamylasemia after ERCP is trained to assist doctors in risk assessment and management of postoperative patients.

[0137] As an implementation manner, step S200, determining the monitoring index characterization vector corresponding to each ERCP post-operative monitoring data sample in the ERCP post-operative monitoring data sample library, may specifically include:

[0138] Step S210: Retrieve the target representation vector mining model, perform representation vector mining on each ERCP postoperative monitoring data sample in the ERCP postoperative monitoring data sample library, and obtain the monitoring index representation vector. The target representation vector mining model is trained by the following steps:

[0139] Step S21: Acquire monitoring indicator data corresponding to a plurality of post-ERCP monitoring data samples, wherein the post-ERCP monitoring data samples include a plurality of monitoring indicator data.

[0140] In an embodiment of the present application, step S21 obtains monitoring index data of multiple samples from a sample library of postoperative ERCP monitoring data. These monitoring index data are the basis of model training and reflect the status and risk of postoperative recovery of patients. In this step, the computer system accesses the sample library of postoperative ERCP monitoring data, which is a collection of a large number of postoperative ERCP patient monitoring data. Each monitoring data sample corresponds to a specific patient and contains multiple monitoring index data of the patient at different time points after surgery. These monitoring index data generally cover multiple aspects such as serum amylase levels, abdominal pain scores, and operation duration, which together constitute a comprehensive data set describing the patient's postoperative recovery status.

[0141] Specifically, the computer system randomly selects a certain number of ERCP postoperative monitoring data samples from the sample library. For each sample, the computer system extracts the corresponding monitoring indicator data. For example, for a certain sample, the computer system can extract the serum amylase level data 1 hour, 6 hours and 24 hours after surgery, as well as the abdominal pain score data at different time points after surgery. These data will be used as input for the subsequent training of the target representation vector mining model.

[0142] Through the processing of step S21, the computer system can obtain a large number of ERCP postoperative monitoring data samples containing multiple monitoring index data. These data samples provide a rich data source for the subsequent training of the target representation vector mining model, so that the model can learn the ability to extract key features from the original monitoring data, providing strong support for the prediction of hyperamylasemia after ERCP.

[0143] Step S22: Based on the representation vector mining model, representation vector mining is performed on the monitoring indicator data corresponding to multiple post-ERCP monitoring data samples, and the monitoring indicator representation vectors obtained by mining are projected into the representation vector domain to obtain the vector positions of the monitoring indicator representation vectors corresponding to the multiple post-ERCP monitoring data samples in the representation vector domain.

[0144] In the embodiment of the present application, step S22 is a key data processing step, which involves using a characterization vector mining model to perform in-depth analysis and conversion of the monitoring index data in the ERCP postoperative monitoring data sample. The purpose of this step is to convert the original monitoring index data into monitoring index characterization vectors that are easier for the machine learning model to understand and process, and map these vectors to a specific characterization vector domain (i.e., a feature space). First, the computer system loads a pre-trained characterization vector mining model. This model may be constructed based on an algorithm of deep learning or machine learning, such as an autoencoder, principal component analysis (PCA) or variational autoencoder (VAE), etc. These models have powerful feature extraction and characterization learning capabilities, and can extract key features from complex raw data and generate corresponding characterization vectors.

[0145] Next, the computer system will traverse each sample in the ERCP postoperative monitoring data sample library and extract the monitoring indicator data corresponding to each sample. These monitoring indicator data usually include multiple dimensions, such as serum amylase level, abdominal pain score, operation time, etc., which together describe the patient's postoperative recovery status and risks.

[0146] Then, the computer system inputs the monitoring indicator data of each sample into the representation vector mining model. The model will perform complex nonlinear transformations on the data, extract key feature information, and generate a fixed-length numerical vector as the representation vector. This representation vector contains the core features of the original data and removes redundant and noisy information, making it easier for subsequent machine learning models to learn useful patterns from it.

[0147] The generated monitoring indicator representation vectors are then projected into a representation vector domain. This representation vector domain is a high-dimensional space in which each point represents a specific monitoring indicator representation vector. By mapping the representation vectors into this space, computer systems can more easily analyze and process them. In this space, representation vectors with similar characteristics will gather together to form clusters, while representation vectors with different characteristics will be distributed in different areas.

[0148] For example, suppose there is a sample of post-ERCP monitoring data, whose monitoring indicator data include serum amylase levels and post-operative abdominal pain scores 1 hour, 6 hours and 24 hours after surgery. The computer system inputs this data into the representation vector mining model, and the model generates a fixed-length representation vector (such as a 10-dimensional vector). Then, this representation vector is projected into a representation vector domain to obtain a specific vector position. In this domain, the representation vectors of other samples with similar serum amylase levels and abdominal pain scores will be relatively close, forming a cluster.

[0149] Through the processing of step S22, the computer system can convert the original post-ERCP monitoring data into a monitoring index representation vector that is easier to understand and process, and provide strong data support for subsequent model training and analysis.

[0150] Step S23: Based on the vector positions of the monitoring indicator representation vectors corresponding to the multiple ERCP post-operative monitoring data samples in the representation vector domain, cluster analysis is performed on the monitoring indicator representation vectors corresponding to the multiple ERCP post-operative monitoring data samples to obtain multiple classification clusters, which include multiple monitoring indicator representation vectors that meet the vector space similarity.

[0151] In the embodiment of the present application, step S23 is a key data analysis step, which performs cluster analysis on these vectors based on the position of the monitoring index characterization vector in the characterization vector domain to identify the inherent structure and pattern in the data. In step S23, the computer system obtains the monitoring index characterization vectors corresponding to the multiple ERCP postoperative monitoring data samples generated in step S22. These characterization vectors have been mapped to a characterization vector domain (i.e., feature space), and each vector has a specific position in this domain. This position represents the coordinates of the vector in the feature space, and also reflects the distribution of the monitoring data samples represented by it on the features. Then, the computer system uses a cluster analysis algorithm to group these monitoring index characterization vectors. The goal of the clustering algorithm is to cluster similar vectors together to form different classification clusters. In this scenario, commonly used clustering algorithms include K-means, hierarchical clustering, DBSCAN, etc. These algorithms can automatically divide the data into several non-intersecting subsets (i.e., cluster clusters) according to the distance or similarity between the vectors.

[0152] Taking the K-means algorithm as an example, the computer system randomly selects K vectors as the initial cluster centers, then calculates the distance from each monitoring indicator representation vector to each cluster center, and assigns the vector to the cluster corresponding to the nearest cluster center. After completing the assignment of all vectors, the algorithm recalculates the mean of all vectors in each cluster and uses this mean as the new cluster center. This process will be iterated continuously until the cluster center no longer changes significantly or the preset number of iterations is reached. Finally, the computer system obtains K classification clusters, each of which contains monitoring indicator representation vectors that meet the vector space similarity (such as the distance is less than the distance threshold).

[0153] These classification clusters represent different postoperative recovery states or risk levels. For example, the vectors in one cluster may represent patients who have recovered well after surgery, and their monitoring indicator representation vectors are clustered together in the feature space to form tight clusters; while the vectors in another cluster may represent patients who are at risk of postoperative hyperamylasemia, and their vectors tend to be dispersed or away from other clusters in the feature space.

[0154] Through the cluster analysis in step S23, the computer system can reveal the inherent structure and pattern in the post-ERCP monitoring data, providing important data support for subsequent target grouping and model training. This method of clustering based on the vector position in the representation vector domain can more accurately capture the key features in the data and improve the accuracy and generalization ability of the prediction model.

[0155] Step S24: performing monitoring indicator occurrence rate analysis on the monitoring indicator data corresponding to the monitoring indicator representation vectors in the same classification cluster to obtain the cluster core indicators corresponding to the classification cluster.

[0156] In step S24, the computer system reviews the multiple classification clusters obtained by the clustering algorithm in step S23. Each cluster contains a group of ERCP postoperative monitoring data samples that are similar in the monitoring indicator representation vector, and these samples have common characteristics in the monitoring indicators. Then, for each classification cluster, the computer system extracts the original monitoring indicator data corresponding to all monitoring indicator representation vectors in the cluster. These original data contain information in multiple dimensions such as serum amylase level, abdominal pain score, and operation time, which together describe the patient's postoperative recovery status and risks. Then, the computer system performs an occurrence rate analysis on the monitoring indicator data in each cluster. Specifically, the computer system counts the number of times or frequency that each monitoring indicator appears in each cluster, and calculates its occurrence rate. Monitoring indicators with a high occurrence rate are more representative in the cluster and may be the key features of the postoperative recovery status or risk level represented by the cluster.

[0157] For example, suppose there is a classification cluster representing a group of patients who have recovered well after surgery. By analyzing the occurrence rate of the monitoring indicator data in this cluster, the computer system may find that the occurrence rates of serum amylase level and abdominal pain score are low, and the values ​​are stable within the normal range. This shows that these two indicators are important for describing the state of good recovery after surgery, and therefore can be identified as the core indicators of this cluster.

[0158] Similarly, for other classification clusters representing different postoperative recovery states or risk levels, the computer system will also perform similar monitoring indicator occurrence rate analysis to identify their respective core indicators. These core indicators not only help to understand the characteristics of different clusters, but also provide targeted feature selection basis for subsequent target grouping determination and model training. Through the monitoring indicator occurrence rate analysis of step S24, the computer system can accurately identify the core monitoring indicators of each classification cluster. These core indicators provide valuable information for subsequent model training and optimization, which helps to improve the prediction accuracy and generalization ability of the model.

[0159] Step S25: Based on the cluster core indicators and the monitoring indicator feature mining focus of the characterization vector mining model, the characterization vector mining model is trained until the characterization vector mining model reaches a preset convergence condition to obtain a trained characterization vector mining model.

[0160] In the embodiment of the present application, step S25 is an iterative optimization step, which further trains and adjusts the characterization vector mining model based on the cluster core indicators and the monitoring indicator feature mining focus of the characterization vector mining model determined in the previous step. The purpose of this step is to optimize the model so that it can better capture the key features in the data and improve the accuracy and generalization ability of the subsequent prediction model.

[0161] In step S25, the computer system reviews the core indicators of each classification cluster determined in step S24. These core indicators represent the key features of different clusters and are an important basis for distinguishing different postoperative recovery states or risk levels. At the same time, the computer system will also consider the monitoring indicator feature mining focus set by the representation vector mining model during the training process. These focuses are usually determined based on domain knowledge and actual application needs, and are used to guide the model to focus on specific monitoring indicators during feature extraction and representation learning.

[0162] Next, the computer system trains the representation vector mining model based on the key points of cluster core indicators and monitoring indicator feature mining. The training process usually includes two steps: forward propagation and back propagation. In the forward propagation stage, the model converts the input data (i.e., monitoring indicator data) into a monitoring indicator representation vector and outputs the prediction results in a specific way (such as a classifier or a regressor). Then, the computer system calculates the loss function value (such as based on the cross entropy loss function or the mean square error loss function) based on the true label and the prediction result to measure the prediction performance of the model. In the back propagation stage, the computer system updates the parameters of the model according to the loss function value. Specifically, the computer system uses optimization algorithms such as gradient descent to calculate the gradient of the loss function with respect to the model parameters, and updates the parameter values ​​according to the gradient information. This process will be iterated continuously until the model reaches the preset convergence condition. The convergence condition can be that the loss function value drops below a certain threshold, the number of iterations reaches a preset upper limit, or the performance of the model on the validation set is no longer significantly improved.

[0163] For example, assume that in step S24, it is determined that the core indicators of a certain classification cluster are serum amylase levels and abdominal pain scores. At the same time, the focus of monitoring indicator feature mining set by the representation vector mining model during the training process also includes these two indicators. In step S25, the computer system trains the model based on this information and input data. Through continuous iterative optimization, the model will gradually learn how to better extract key features related to serum amylase levels and abdominal pain scores from the input data, and generate more representative monitoring indicator representation vectors. Finally, when the model reaches the preset convergence condition, the training process ends, and an optimized representation vector mining model is obtained.

[0164] Through the training process of step S25, the computer system can obtain a more accurate and effective representation vector mining model. This model can better capture the key features in the data and provide more valuable input data for the subsequent post-ERCP hyperamylasemia prediction model. At the same time, by continuously optimizing the model parameters and structure, the computer system can also improve the generalization ability of the model, so that it can better adapt to new data and environmental changes.

[0165] As an implementation manner, step S24 performs monitoring indicator occurrence rate analysis on monitoring indicator data corresponding to monitoring indicator representation vectors in the same classification cluster to obtain cluster core indicators corresponding to the classification cluster, which may specifically include:

[0166] Step S241: performing monitoring indicator occurrence rate analysis on the monitoring indicator data corresponding to the monitoring indicator representation vectors in the same classification cluster to obtain the occurrence rate of each monitoring indicator data.

[0167] In an embodiment of the present application, step S241 is a key step in analyzing the monitoring indicator data in each classification cluster. The purpose of this step is to determine the occurrence rate of each monitoring indicator in the cluster by counting the number of times or frequency that each monitoring indicator appears in the same classification cluster, thereby providing data support for the subsequent identification of cluster core indicators. Specifically, in step S241, the computer system focuses on each classification cluster obtained by the clustering algorithm. Each cluster represents a group of ERCP postoperative monitoring data samples that are similar in monitoring indicators. Then, for each cluster, the computer system extracts the original monitoring indicator data corresponding to all monitoring indicator characterization vectors in the cluster. These data typically include monitoring indicators of multiple dimensions, such as serum amylase levels, abdominal pain scores, operation duration, etc.

[0168] Then, the computer system performs an occurrence analysis on these monitoring indicator data. This analysis process is mainly completed by counting the number of times or frequencies of each monitoring indicator in the cluster. For example, for the indicator of serum amylase level, the computer system counts the number of times it appears in all samples in a cluster and calculates its occurrence rate. The occurrence rate can be obtained by dividing the number of occurrences by the total number of samples in the cluster, which reflects the importance or representativeness of the indicator in the cluster.

[0169] Through the occurrence rate analysis in step S241, the computer system can generate a set of occurrence rate data for each classification cluster, which will be used to identify the core indicators of the cluster in the subsequent steps, thereby helping to understand the postoperative recovery status or risk level represented by different clusters. This data-driven analysis method helps to improve the accuracy and generalization ability of the prediction model and provide doctors with a more reliable postoperative risk assessment tool.

[0170] Step S242: When the appearance rate of the target monitoring indicator data reaches a preset appearance rate, the target monitoring indicator data is determined as the cluster core indicator corresponding to the classification cluster.

[0171] In step S242, the computer system reviews the occurrence rate of each monitoring indicator calculated in step S241 in the classification cluster. These occurrence rates are obtained by counting the number of times or frequencies of each indicator appearing in the cluster, and they reflect the importance or representativeness of the indicator in the cluster. Then, the computer system sets a preset occurrence rate threshold. This threshold is usually determined based on actual application requirements and data distribution, and is used to screen out monitoring indicators with sufficiently high occurrence rates that are sufficient to represent cluster characteristics. The setting of the preset occurrence rate threshold is a balancing process, and it is necessary to ensure that the screened indicators are both representative enough and not too broad or contain too much redundant information. Then, the computer system compares the occurrence rate of each monitoring indicator with the preset occurrence rate threshold one by one. For those monitoring indicators whose occurrence rates reach or exceed the threshold, the computer system identifies them as target monitoring indicators and determines them as cluster core indicators of the corresponding classification cluster. These core indicators represent the main characteristics of the cluster and are an important basis for subsequent model training and analysis.

[0172] For example, suppose there is a classification cluster representing a group of patients who recover well after surgery. In step S241, the computing system has calculated the occurrence rate of each monitoring indicator in the cluster. Assuming that the preset occurrence rate threshold is set to 80%, then in step S242, the computer system compares the occurrence rate of each indicator with this threshold. If the occurrence rate of an indicator (such as abdominal pain score) reaches 80% or higher, then the indicator will be determined as one of the cluster core indicators of the cluster. This means that the abdominal pain score is a particularly important or representative monitoring indicator for this cluster representing a good state of recovery after surgery.

[0173] Through the processing of step S242, the computer system can accurately identify the cluster core indicators of each classification cluster. These core indicators not only help to understand the postoperative recovery status or risk level represented by different clusters, but also provide a targeted feature selection basis for subsequent model training. This helps to improve the accuracy and generalization ability of the prediction model, so that the model can more accurately predict the risk of hyperamylasemia after ERCP.

[0174] As an implementation mode, step S25, training the representation vector mining model according to the cluster core indicators and the monitoring indicator feature mining focus of the representation vector mining model, may specifically include:

[0175] Step S25a: When the cluster core indicators do not correspond to the monitoring indicator feature mining focus corresponding to the representation vector mining model, the error function of the training representation vector mining model is optimized according to the monitoring indicator feature mining focus; the representation vector mining model is trained based on the optimized error function.

[0176] When it is found that the model fails to pay sufficient attention to the features closely related to the core indicators of the cluster during feature mining, the computer system guides the model to better learn these key features by optimizing the error function. Specifically, in step S25a, the computer system analyzes the focus of monitoring indicator feature mining of the current characterization vector mining model and compares the core indicators of the cluster. If it is found that there is a significant difference between the two, that is, the model fails to pay sufficient attention to the features closely related to the core indicators of the cluster during the feature mining process, then the computer system will take corresponding optimization measures. The key to optimization is to adjust the error function (or loss function), which is a core component in the training process of the machine learning model and is used to measure the difference between the model prediction results and the true label. By adjusting the weights related to the core indicators of the cluster in the error function, the computer system can guide the model to pay more attention to these key features during the training process.

[0177] For example, suppose that the core indicators of the cluster mainly include serum amylase level and abdominal pain score, but the current representation vector mining model pays more attention to other indicators such as operation duration when mining features. In this case, the computer system can adjust the weights related to serum amylase level and abdominal pain score in the error function to give them higher priority. In this way, during the model training process, once there is a difference between the predicted result and the true label in serum amylase level or abdominal pain score, the error function will produce a larger penalty, thereby prompting the model to pay more attention to these two key features. Based on the optimized error function, the computer system retrains the representation vector mining model. Through the iterative optimization process, the model will gradually learn the key features that are closely related to the core indicators of the cluster, and show higher accuracy and robustness in subsequent prediction tasks.

[0178] In this way, step S25a effectively solves the problem that the core indicators of the cluster do not correspond to the key points of the feature mining of the monitoring indicators, and improves the training effect and prediction performance of the representation vector mining model. In the scenario of training the prediction model of hyperamylasemia after ERCP, this optimization strategy is of great significance for improving the accuracy and reliability of the model.

[0179] Or; Step S25b: When the cluster core indicators do not correspond to the monitoring indicator feature mining focus corresponding to the representation vector mining model, the ERCP postoperative monitoring data sample set is obtained based on the monitoring indicator feature mining focus, and the ERCP postoperative monitoring data samples in the ERCP postoperative monitoring data sample set are samples corresponding to the monitoring indicator feature mining focus; the representation vector mining model is trained based on the ERCP postoperative monitoring data samples in the ERCP postoperative monitoring data sample set.

[0180] In the embodiment of the present application, step S25b is intended to solve the problem that the core index of the cluster does not correspond to the key point of the monitoring index feature mining of the characterization vector mining model. When the computer system identifies that the key point of the monitoring index feature mining of the current model is inconsistent with the core index of the cluster, a more satisfactory ERCP postoperative monitoring data sample set is reacquired according to the key point of the monitoring index feature mining, and the characterization vector mining model is trained based on this new sample set. Specifically, in step S25b, the computer system reviews the key point of the monitoring index feature mining of the current characterization vector mining model, which is set or automatically learned by the model during the training process, and is used to guide the model to pay attention to which monitoring indicators during feature extraction. Then, the computer system compares this key point with the core index of the cluster determined by step S24. If it is found that there is a significant inconsistency between the two, that is, the monitoring index currently concerned by the model is not the core index of the cluster, then it is necessary to adjust the training data sample set. In order to obtain a more satisfactory ERCP postoperative monitoring data sample set that meets the key point of the monitoring index feature mining, the computer system selects samples containing monitoring index data corresponding to the key point of the monitoring index feature mining from the original ERCP postoperative monitoring data sample library. These screened samples will form a new training data sample set, which is characterized by being richer and more comprehensive in monitoring indicators, especially including those key indicator data that the model currently focuses on but may lack. For example, suppose that the current monitoring indicator feature mining focuses on serum amylase levels and abdominal pain scores, but the data of these two indicators in the original training data sample set are not rich or representative enough. In this case, the computer system screens out those samples with complete and representative data of serum amylase levels and abdominal pain scores from the sample library to form a new training data sample set. Then, the representation vector mining model is trained based on this new sample set to ensure that the model can fully pay attention to these key indicators during feature extraction and representation learning.

[0181] Through the implementation of step S25b, the computer system can effectively adjust the training data sample set to make it more consistent with the model's monitoring indicator feature mining focus. This can not only improve the model's training effect, but also ensure that the model can accurately identify key features closely related to the occurrence of hyperamylasemia in subsequent prediction tasks, thereby improving the accuracy and reliability of the prediction. In the scenario of training a prediction model for hyperamylasemia after ERCP, this data sample set adjustment strategy is of great significance for improving the performance of the model.

[0182] Or; Step S25c: When the cluster core indicators do not correspond to the monitoring indicator feature mining focus corresponding to the representation vector mining model, when obtaining the monitoring indicator data corresponding to multiple ERCP post-operative monitoring data samples, specific monitoring indicator data corresponding to the monitoring indicator feature mining focus is added to the monitoring indicator data.

[0183] In the embodiment of the present application, step S25c provides a flexible data enhancement strategy for solving the problem that the core indicators of the cluster do not correspond to the monitoring indicator feature mining focus of the characterization vector mining model. When the computer system identifies that the monitoring indicator feature mining focus of the current model is inconsistent with the core indicators of the cluster, it will add specific monitoring indicator data corresponding to the monitoring indicator feature mining focus in a targeted manner during the process of acquiring the monitoring indicator data to enrich and enhance the training data set.

[0184] Specifically, in step S25c, the computer system analyzes the monitoring indicator feature mining focus of the current characterization vector mining model, which is the key feature set or automatically learned by the model during the training process. Then, the computer system compares these key points with the cluster core indicators determined by step S24. If it is found that there is a significant inconsistency between the two, that is, the monitoring indicator currently concerned by the model does not completely match the cluster core indicator, then it is necessary to adjust it in the data acquisition stage. In order to enhance the training data set, the computer system deliberately adds specific monitoring indicator data corresponding to the monitoring indicator feature mining focus when acquiring the monitoring indicator data corresponding to multiple ERCP postoperative monitoring data samples. These specific data can be indicator data that was not previously included in the training data set, or are relatively scarce but of great significance in the existing data set. By adding these specific data, the computer system can ensure that the training data set is richer and more comprehensive in monitoring indicators, thereby better covering the features that the model is concerned about.

[0185] For example, suppose that the current model focuses on the monitoring indicator feature mining including serum amylase level and C-reactive protein level, but the data of C-reactive protein level in the existing training data set is relatively scarce. In this case, when the computer system obtains new post-ERCP monitoring data samples, it pays special attention to and adds the monitoring data of C-reactive protein level. These newly added data will form a more complete and balanced training data set together with the existing data such as serum amylase level.

[0186] Through the implementation of step S25c, the computer system can flexibly adjust the training data set to make it more consistent with the model's monitoring indicator feature mining focus. This data enhancement strategy can not only improve the model's training effect, but also ensure that the model can accurately identify key features closely related to the occurrence of hyperamylasemia in subsequent prediction tasks. In the scenario of training a prediction model for hyperamylasemia after ERCP, this strategy is of great significance for improving the model's prediction accuracy and generalization ability.

[0187] As an implementation mode, step S23, based on the vector positions of the monitoring indicator characterization vectors corresponding to the multiple ERCP post-operative monitoring data samples in the characterization vector domain, cluster analysis is performed on the monitoring indicator characterization vectors corresponding to the multiple ERCP post-operative monitoring data samples to obtain multiple classification clusters, which may specifically include:

[0188] Step S231: determining the spatial similarity between the monitoring indicator representation vectors according to the distribution positions of the monitoring indicator representation vectors corresponding to the plurality of post-ERCP monitoring data samples in the representation vector domain.

[0189] In the embodiment of the present application, step S231 is an important data processing step, which involves analyzing the distribution positions of monitoring indicator characterization vectors corresponding to multiple ERCP post-operative monitoring data samples in the characterization vector domain to determine the spatial similarity between these vectors.

[0190] In step S231, the computer system accesses the characterization vector domain, which is a high-dimensional space, and each monitoring indicator characterization vector is a point in this space. These vectors are extracted and converted from the original post-ERCP monitoring data sample through the previous steps (such as step S22), and they contain the key feature information of the sample on multiple monitoring indicators. Next, the computer system analyzes the distribution position of these monitoring indicator characterization vectors in the characterization vector domain. Since the characterization vector domain is a high-dimensional space, the computer system uses an appropriate distance measurement method to calculate the distance between vectors. Commonly used distance measurement methods include Euclidean distance, cosine similarity, etc. These distance measurement methods can quantify the proximity of vectors in space, thereby reflecting the similarity between vectors.

[0191] For example, suppose there are two monitoring indicator representation vectors V1 and V2, which represent two different ERCP postoperative monitoring data samples. The computer system uses Euclidean distance to calculate the distance between the two vectors. In addition to Euclidean distance, the computer system can also use other distance measurement methods such as cosine similarity to calculate the similarity between vectors. Cosine similarity measures the similarity of two vectors by calculating the cosine value of the angle between them. The value range is [-1,1], where 1 means completely similar and -1 means completely opposite.

[0192] Through the processing of step S231, the computer system can determine the spatial similarity between the monitoring indicator representation vectors corresponding to multiple ERCP postoperative monitoring data samples. This similarity information will be used in the subsequent cluster analysis step (such as step S232) to help the computer system divide the data samples into different classification clusters, thereby revealing the inherent structure and pattern in the data. These classification clusters are of great significance for understanding postoperative recovery status, identifying high-risk patient groups, and training more accurate prediction models.

[0193] Step S232: when the spatial similarity is greater than the spatial similarity threshold, it is determined that the monitoring indicator representation vectors whose spatial similarity is less than the spatial similarity threshold are located in the same classification cluster.

[0194] In the embodiment of the present application, step S232 is a key clustering decision step, which determines which vectors should be classified into the same classification cluster based on the spatial similarity between the monitoring indicator representation vectors. This step is of great significance for revealing the inherent structure and pattern in the data, as well as subsequent target grouping and model training. In step S232, the computer system reviews the spatial similarity between the monitoring indicator representation vectors calculated in step S231. These similarities are calculated by appropriate distance metrics (such as Euclidean distance, cosine similarity, etc.), which quantify the proximity of the vectors in the representation vector domain (feature space). Then, the computer system sets a spatial similarity threshold. This threshold is a preset parameter for determining which vectors are close enough in space to be classified into the same classification cluster. The choice of threshold is usually based on factors such as the actual distribution of data, the purpose of clustering, and domain knowledge.

[0195] Then, the computer system compares the spatial similarity between each pair of monitoring indicator representation vectors one by one, and compares them with the spatial similarity threshold. If the similarity between two vectors is greater than or equal to the threshold, then the two vectors are considered to be similar and should be classified into the same classification cluster. For example, suppose there are two monitoring indicator representation vectors V1 and V2, which represent two different ERCP postoperative monitoring data samples. In step S231, the computer system calculates that the spatial similarity between V1 and V2 is 0.8 (assuming that cosine similarity is used as the measurement method, the value range is [-1,1], where 1 represents complete similarity). If the preset spatial similarity threshold is 0.7, then since 0.8 is greater than 0.7, the computer system will classify V1 and V2 into the same classification cluster.

[0196] Through this process, the computer system is able to classify similar vectors into the same classification cluster based on the spatial similarity between the monitoring indicator representation vectors. These classification clusters represent different subsets in the data, each of which has similar monitoring indicator characteristics. These classification clusters provide an important data basis for subsequent target grouping and model training, which helps to reveal the patterns of postoperative recovery status, identify high-risk patient groups, and improve the accuracy of prediction models.

[0197] As an implementation mode, step S20, based on the data item sequences corresponding to the samples of the non-hyperamylasemia group and the samples of the hyperamylasemia group, respectively, determines the data item commonality measurement value of the samples of the non-hyperamylasemia group and the samples of the hyperamylasemia group, which may specifically include:

[0198] Step S201: determining a target overlapping data sequence of the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia according to the data item sequences corresponding to the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia respectively;

[0199] Step S202: determining the length of the data item sequences corresponding to the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia respectively;

[0200] Step S203: determining a first data item commonality metric value between the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group based on the data item sequence lengths and the target overlapping data sequence corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group respectively;

[0201] Step S204: based on the first data item commonality metric value, determine the data item commonality metric value between the hyperamylasemia non-occurring group samples and the hyperamylasemia occurring group samples.

[0202] In the embodiment of the present application, step S201 is to find the overlap between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia in the data item sequence. This step is of great significance for understanding the common features of the two groups of samples in time series data, and provides a basis for subsequent data analysis and model training.

[0203] Specifically, in step S201, the computer system obtains data item sequences corresponding to samples of the group without hyperamylasemia and samples of the group with hyperamylasemia. These data item sequences usually contain monitoring data at different time points after surgery, such as serum amylase levels, abdominal pain scores, etc. They are arranged in chronological order to form a time series data set that can reflect the patient's postoperative recovery status or risk changes.

[0204] Then, the computer system compares the two groups of data item sequences one by one to find the overlapping parts between them. This process can be implemented by a variety of algorithms, such as the LCS algorithm (Longest Common Subsequence, LCS). The LCS algorithm can find the longest subsequence in the two sequences so that this subsequence appears in the same relative order in both sequences. In the embodiment of the present application, the LCS algorithm can help the computer system find the overlapping parts in the two groups of sample data item sequences, that is, the monitoring data items they share. For example, assuming that the data item sequence of the sample A in the group with no hyperamylasemia is [T1: serum amylase level 150U / L, T2: abdominal pain score 2, T3: serum amylase level 140U / L], and the data item sequence of the sample B in the group with hyperamylasemia is [T1: serum amylase level 160U / L, T2: abdominal pain score 3, T4: operation time 60 minutes]. By comparing the two sequences through the LCS algorithm, the computer system can find that their data items at time points T1 and T2 are overlapped, that is, the two monitoring indicators of serum amylase level and abdominal pain score are recorded in both samples. Therefore, this part of the overlapping data item sequence is the target overlapping data sequence to be determined in step S201.

[0205] Through the processing of step S201, the computer system can accurately find the overlapped parts of the data item sequences of the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia. These overlapping data item sequences not only reflect the common characteristics of the two groups of samples in time series data, but also provide valuable information for subsequent data analysis and model training.

[0206] In step S202, it is involved to determine the length of the data item sequence corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia. The purpose of this step is to quantitatively evaluate the coverage and time span of the two groups of samples on the time series data, and provide important reference information for subsequent data analysis and model training. Specifically, in step S202, the computer system reviews the data item sequences of the samples of the group without hyperamylasemia and the group with hyperamylasemia determined in step S201. These data item sequences contain monitoring data at different time points after surgery, such as serum amylase levels, abdominal pain scores, etc. They are arranged in chronological order to form a time series data set that can reflect the patient's postoperative recovery status or risk changes. Then, the computer system calculates the length of the data item sequences of the two groups of samples respectively. The sequence length refers to the number of monitoring data points contained in the data item sequence, which reflects the coverage and time span of the sample on the time series data. For each sample, the computer system counts the number of monitoring data points in its data item sequence to obtain the sequence length of the sample.

[0207] For example, if the data item sequence of sample A in the group without hyperamylasemia is [T1: serum amylase level 150U / L, T2: abdominal pain score 2, T3: serum amylase level 140U / L], then the length of this sequence is 3, because it contains 3 monitoring data points. Similarly, if the data item sequence of sample B in the group with hyperamylasemia is [T1: serum amylase level 160U / L, T2: abdominal pain score 3, T4: operation time 60 minutes, T5: serum amylase level 180U / L], then the length of this sequence is 4.

[0208] Through the processing of step S202, the computer system can accurately calculate the length of the data item sequence corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia. These length information not only reflects the coverage and time span of the two groups of samples on the time series data, but also provides an important reference for subsequent data analysis and model training. For example, in the subsequent data grouping or model training process, the computer system can screen or weight the samples according to the sequence length to better utilize the information in the time series data.

[0209] In step S203, based on the data item sequence length of the hyperamylasemia non-occurrence group samples and the hyperamylasemia occurrence group samples and the target overlapping data sequence, the first data item commonality metric between the two groups of samples is determined. The purpose of this step is to quantitatively evaluate the similarity or difference between the two groups of samples in the time series data, and provide an important basis for subsequent data processing and model training. Specifically, in step S203, the computer system reviews the results obtained in steps S201 and S202. Step S201 determines the target overlapping data sequence of the hyperamylasemia non-occurrence group samples and the hyperamylasemia occurrence group samples, that is, the common part of the two groups of samples in the time series data. Step S202 calculates the data item sequence length of the two groups of samples, reflecting their coverage and time span in the time series data. Next, the computer system will combine these two information to determine the first data item commonality metric. This metric is intended to quantitatively evaluate the similarity or difference between the two groups of samples in the overlapping data items, as well as their relationship in the coverage and time span of the time series data. The specific calculation method may involve various factors, such as the number of overlapping data items, numerical differences, changing trends, and the ratio of the lengths of data item sequences.

[0210] For example, suppose that the data item sequence length of sample A in the group without hyperamylasemia is 3, and the target overlapping data sequence contains data from two time points (T1 and T2); while the data item sequence length of sample B in the group with hyperamylasemia is 4, and also contains the data from these two time points (T1 and T2) as the target overlapping part. In this case, the computer system can calculate the ratio of the number of overlapping data items to the length of each sequence, i.e., 2 / 3 and 2 / 4, as an indicator for evaluating the similarity of the two groups of samples in the overlapping data items. In addition, the computer system can further analyze the numerical differences and changing trends of the overlapping data items to obtain a more comprehensive similarity assessment.

[0211] Through the processing of step S203, the computer system can obtain a quantitative commonality measurement value of the first data item, which reflects the similarity or difference between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia in the time series data. This measurement value will provide an important reference for subsequent steps such as data grouping, feature selection and model training, and help improve the accuracy and generalization ability of the prediction model.

[0212] In step S204, based on the first data item commonality metric value calculated previously, the data item commonality metric value between the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia is further determined. This metric value comprehensively reflects the overall similarity or difference between the two groups of samples in the time series data, which is of great significance for subsequent data grouping, feature selection and model training.

[0213] Specifically, in step S204, the computer system reviews the first data item commonality metric calculated in step S203. This value is usually calculated based on multiple factors such as the number of overlapping data items, numerical differences, change trends, and the ratio of data item sequence lengths, and is used to quantitatively evaluate the similarity or difference between two groups of samples on overlapping data items. Then, the computer system determines the final data item commonality metric based on this first data item commonality metric and other possible factors (such as domain knowledge, data distribution characteristics, etc.). This metric not only takes into account the information of overlapping data items, but also may incorporate the characteristics of other non-overlapping data items to more comprehensively reflect the overall similarity or difference between the two groups of samples on time series data. In the process of determining the data item commonality metric, the computer system can adopt some strategies or methods. For example, it can set a weight for the first data item commonality metric, and perform weighted calculation in combination with other features or indicators to obtain the final metric. In addition, the computer system can also adjust and optimize the calculation process of the metric according to the actual situation of the data and the needs of the prediction model.

[0214] For example, assuming that the commonality metric value of the first data item calculated in step S203 is 0.8 (indicating that the two groups of samples have a high similarity in the overlapping data items). In step S204, the computer system may consider the characteristics of other non-overlapping data items, such as the trend of serum amylase levels at different time points after surgery, the fluctuation of abdominal pain scores, etc. By comprehensively analyzing these factors, the computer system ultimately determines that the commonality metric value of the data item is 0.75, which reflects both the similarity in the overlapping data items and the characteristic differences of other non-overlapping data items.

[0215] Through the processing of step S204, the computer system can obtain a quantitative data item commonality measurement value, which comprehensively reflects the overall similarity or difference between the samples of the group without hyperamylasemia and the group with hyperamylasemia in the time series data. This measurement value will provide an important reference for subsequent data grouping, feature selection and model training, and help improve the accuracy and generalization ability of the prediction model. At the same time, it also provides doctors with a more in-depth basis for postoperative recovery status assessment and risk prediction, which helps to formulate a more personalized treatment plan.

[0216] As an implementation manner, step S204, based on the commonality metric value of the first data item, determines the commonality metric value of the data item between the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia, which may specifically include:

[0217] Step S2041: determining the overlapping data items of the non-hyperamylasemia group samples and the hyperamylasemia group samples according to the data item sequences corresponding to the non-hyperamylasemia group samples and the hyperamylasemia group samples.

[0218] In the embodiment of the present application, step S2041 is to determine the overlapping data items between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia. These data items are the common features of the two groups of samples in the time series data, which are of great significance for understanding the similarities and differences between the two groups of samples.

[0219] Specifically, in step S2041, the computer system reviews the data item sequences corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia. These data item sequences contain monitoring data at different time points after surgery, such as serum amylase levels, abdominal pain scores, etc. They are arranged in chronological order to form a time series data set that can reflect the patient's postoperative recovery status or risk changes. Then, the computer system compares the two sets of data item sequences one by one to find out the overlapping parts between them. This process can be implemented by programming, by traversing each data item in the two sets of sequences and checking whether they exist in the other set of sequences. If so, these data items are considered to overlap.

[0220] For example, suppose the data item sequence of sample A in the group without hyperamylasemia is [T1: serum amylase level 150U / L, T2: abdominal pain score 2, T3: serum amylase level 140U / L], and the data item sequence of sample B in the group with hyperamylasemia is [T1: serum amylase level 160U / L, T2: abdominal pain score 3, T4: operation time 60 minutes]. By comparing these two sequences, the computer system can find that their data items at the T1 and T2 time points are overlapping, that is, the two monitoring indicators of serum amylase level and abdominal pain score are recorded in both samples. Therefore, serum amylase level and abdominal pain score are determined to be overlapping data items.

[0221] Through the processing of step S2041, the computer system can accurately find the overlapping data items between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia. These overlapping data items not only reflect the common characteristics of the two groups of samples in the time series data, but also provide valuable information for subsequent data analysis and model training. For example, in the subsequent data grouping or feature selection process, overlapping data items can be given priority because they represent the common points of the two groups of samples and are more likely to be related to the prediction target.

[0222] Step S2042: Determine the second data item commonality metric value of the hyperamylasemia non-occurring group samples and the hyperamylasemia occurring group samples based on the data item sequence lengths and overlapping data items corresponding to the hyperamylasemia non-occurring group samples and the hyperamylasemia occurring group samples respectively.

[0223] In the embodiment of the present application, step S2042 determines the commonality metric value of the second data item between the two groups of samples based on the data item sequence length and the overlapping data items of the samples of the group without hyperamylasemia and the group with hyperamylasemia. This metric is intended to quantitatively evaluate the similarity of the two groups of samples in terms of the coverage and integrity of the time series data. Specifically, in step S2042, the computer system reviews the overlapping data items determined in step S2041, and the data item sequence lengths of the samples of the group without hyperamylasemia and the group with hyperamylasemia calculated in step S202. The data item sequence length reflects the coverage of the sample on the time series data, while the overlapping data items represent the common features of the two groups of samples on the time series data. Then, the computer system uses these data to calculate the commonality metric value of the second data item. A common method is to compare the ratio of the number of overlapping data items to the length of each sequence. This ratio reflects the relative importance of the overlapping data items in the two groups of samples, so that the similarity of the two groups of samples in terms of the coverage and integrity of the time series data can be quantitatively evaluated. For example, suppose that the data item sequence length of sample A in the group without hyperamylasemia is 3, and there are 2 data items that overlap with sample B in the group with hyperamylasemia. Similarly, the data item sequence length of sample B in the group with hyperamylasemia is 4, and there are also these 2 overlapping data items. Then, for sample A, the ratio of the overlapping data items to its sequence length is 2 / 3; for sample B, this ratio is 2 / 4. These two ratios can be regarded as the coverage and completeness measures of the two groups of samples on the overlapping data items, and can also be used as part of the commonality measure of the second data item.

[0224] In addition, in order to more comprehensively evaluate the similarity of the two groups of samples, the computer system may also consider other factors, such as the difference in the values ​​of the overlapping data items, the trend of change, etc. These factors can be quantitatively evaluated by calculating additional metrics and combined with the above ratio to form a more comprehensive second data item commonality metric. Through the processing of step S2042, the computer system can obtain a quantitative second data item commonality metric, which comprehensively reflects the similarity between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia in terms of time series data coverage and completeness. This metric will provide important reference information for subsequent data analysis and model training, and help improve the accuracy and generalization ability of the prediction model.

[0225] Step S2043: Determine the data item commonality metric value between the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia based on the first data item commonality metric value and the second data item commonality metric value.

[0226] Specifically, in step S2043, the computer system first reviews the results calculated in steps S2041 and S2042. Step S2041 determines the overlapping data items of the two groups of samples, and calculates the first data item commonality metric value based on these overlapping data items, which mainly reflects the similarity or difference between the two groups of samples in the overlapping data items. Step S2042 calculates the second data item commonality metric value based on the data item sequence length and the number of overlapping data items, which mainly measures the similarity between the two groups of samples in terms of time series data coverage and completeness. In step S2043, the computer system comprehensively considers these two metrics to determine the final data item commonality metric value. This process can be implemented in a variety of ways, such as the weighted average method. Assuming that a weight α is assigned to the commonality measure value of the first data item and a weight β (α+β=1) is assigned to the commonality measure value of the second data item, the final commonality measure value of the data item can be calculated by the following formula: data item commonality measure value=α×first data item commonality measure value+β×second data item commonality measure value.

[0227] The weight distribution can be determined according to the actual application scenario and data characteristics. For example, if it is believed that overlapping data items are more important for evaluating the similarity of two groups of samples, α can be set to a larger value; if it is believed that the coverage and completeness of time series data are equally important, α and β can be set to equal values.

[0228] To illustrate with a specific example, assume that the commonality metric value of the first data item is 0.8 (calculated based on the similarity of the overlapping data items), and the commonality metric value of the second data item is 0.7 (calculated based on the length of the data item sequence and the overlapping data items), and that the two metrics have the same importance, so both α and β are set to 0.5. Then the final commonality metric value of the data items is: commonality metric value of data items = 0.5×0.8+0.5×0.7=0.75.

[0229] This value comprehensively reflects the overall similarity or difference between the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia in the time series data. This metric will provide an important reference for subsequent steps such as data grouping, feature selection, and model training, and help improve the accuracy and generalization ability of the prediction model.

[0230] As an implementation manner, step S2043, based on the first data item commonality metric value and the second data item commonality metric value, determines the data item commonality metric value of the hyperamylasemia non-occurring group samples and the hyperamylasemia occurring group samples, which may specifically include:

[0231] Step S20431: Determine the percentages of risk monitoring indicators corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group, respectively, based on the data item sequences corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group.

[0232] In the embodiment of the present application, step S20431 is a key data analysis step, which determines the percentage of risk monitoring indicators corresponding to the two groups of samples based on the data item sequence of the group samples without hyperamylasemia and the group samples with hyperamylasemia. The purpose of this step is to quantitatively evaluate the importance or representativeness of the monitoring indicators closely related to the occurrence of hyperamylasemia in the time series data of the two groups of samples.

[0233] Specifically, in step S20431, the computer system analyzes the data item sequences corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia. These data item sequences contain monitoring data at different time points after surgery, such as serum amylase levels, abdominal pain scores, etc., which reflect the patient's postoperative recovery status and risks. Then, the computer system identifies those monitoring indicators that are closely related to the occurrence of hyperamylasemia based on domain knowledge and historical data. These indicators are called "risk monitoring indicators". These indicators are usually monitoring data items that have been found to be closely related to the risk of postoperative hyperamylasemia in previous studies or clinical practices. Then, for the samples of the group without hyperamylasemia and the group with hyperamylasemia, the computer system counts the number of occurrences or frequencies of the risk monitoring indicators in their data item sequences, and calculates the percentage of these indicators in the total length of their respective data item sequences. This percentage reflects the importance or representativeness of the risk monitoring indicator in the sample, that is, the risk characteristics of the sample related to the occurrence of hyperamylasemia in the time series data. For example, assume that serum amylase levels and abdominal pain scores are identified as risk monitoring indicators. For sample A in the group without hyperamylasemia, the length of its data item sequence is 5, in which the two indicators of serum amylase level and abdominal pain score appear 3 times each. Then, the percentage of the risk monitoring indicator of sample A is (3+3) / 5=1.2 (note that the percentage here exceeds 100% because there are two indicators, but usually the percentage of each indicator is calculated separately or the average is taken). Similarly, for sample B in the group with hyperamylasemia, the computer system calculates its percentage of risk monitoring indicators in the same way.

[0234] Through the processing of step S20431, the computer system can accurately determine the percentages of risk monitoring indicators corresponding to the samples in the group without hyperamylasemia and the group with hyperamylasemia, respectively. These percentage information provides important input data for the subsequent steps S20432 and S20433, which helps to more comprehensively evaluate the overall similarities or differences between the two groups of samples in the time series data.

[0235] Step S20432: Determine the third data item commonality measurement value of the samples in the group without hyperamylasemia and the group with hyperamylasemia according to the percentages of risk monitoring indicators corresponding to the samples in the group without hyperamylasemia and the group with hyperamylasemia, respectively.

[0236] In the embodiment of the present application, step S20432 is a key data analysis step, which is based on the percentage of risk monitoring indicators corresponding to the samples of the group without hyperamylasemia and the group with hyperamylasemia, respectively, to determine the commonality metric value of the third data item between the two groups of samples. This metric is intended to quantitatively evaluate the similarity or difference between the two groups of samples in the risk monitoring indicators, and provide important reference information for subsequent model training. Specifically, in step S20431, the computer system has calculated the percentage of risk monitoring indicators for the samples of the group without hyperamylasemia and the group with hyperamylasemia. These percentages reflect the importance or representativeness of the risk monitoring indicators in the two groups of sample data item sequences, that is, the correlation between these indicators and the risk of postoperative hyperamylasemia. In step S20432, the computer system further analyzes the two groups of percentage data to determine the commonality metric value between them. This commonality metric value can be calculated by a variety of methods, and the specific method depends on the characteristics of the data and the purpose of the analysis.

[0237] One possible method is to calculate the difference or ratio between the two groups of percentages. For example, the computer system can calculate the difference or ratio between the percentage of the risk monitoring indicator of the sample group without hyperamylasemia and the corresponding percentage of the sample group with hyperamylasemia. This difference or ratio reflects the similarity or difference between the two groups of samples in the risk monitoring indicator. If the difference or ratio is small, it means that the two groups of samples have a high similarity in the risk monitoring indicator; otherwise, it means that there is a large difference between them. Another possible method is to use statistical methods to calculate commonality measures. For example, the computer system can calculate the correlation coefficient or covariance between the two groups of percentages to evaluate the linear relationship or correlation between them. These statistics can provide more comprehensive information to help the computer system more accurately evaluate the similarities or differences between the two groups of samples in the risk monitoring indicators.

[0238] For example, suppose the percentage of the risk monitoring index of the group of samples without hyperamylasemia is 80%, while the corresponding percentage of the group of samples with hyperamylasemia is 90%. The computer system can calculate the difference between these two percentages (90%-80%=10%) as part of the commonality metric of the third data item. This difference reflects the degree of difference between the two groups of samples in the risk monitoring index. If the difference is small, it means that the two groups of samples have a high similarity in the risk monitoring index; otherwise, it means that there is a large difference between them.

[0239] Through the processing of step S20432, the computer system can obtain a quantitative third data item commonality measurement value, which comprehensively reflects the similarity or difference between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia in terms of risk monitoring indicators. This measurement value will provide important reference information for subsequent steps such as data grouping, feature selection, and model training, and help improve the accuracy and generalization ability of the prediction model.

[0240] Step S20433: Determine the data item commonality measurement value between the samples in the group without hyperamylasemia and the samples in the group with hyperamylasemia based on the first data item commonality measurement value, the second data item commonality measurement value and the third data item commonality measurement value.

[0241] In the embodiment of the present application, step S20433 determines the data item commonality measurement value between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia based on the first data item commonality measurement value, the second data item commonality measurement value, and the third data item commonality measurement value calculated in the previous step. The purpose of this step is to comprehensively consider multiple measurement values ​​to obtain an indicator that comprehensively reflects the similarity or difference between the two groups of samples in time series data, providing strong support for subsequent model training and data analysis.

[0242] Specifically, in step S20433, the computer system reviews the three data item commonality metrics calculated in the previous steps: the first data item commonality metric is mainly calculated based on the difference and change trend of the overlapping data items, reflecting the similarity of the two groups of samples in the overlapping data items; the second data item commonality metric is calculated based on the length of the data item sequence and the number of overlapping data items, measuring the similarity of the two groups of samples in the coverage and integrity of the time series data; the third data item commonality metric focuses on the percentage of risk monitoring indicators, evaluating the similarity or difference of the two groups of samples in the monitoring indicators closely related to the occurrence of hyperamylasemia. Next, the computer system uses an appropriate comprehensive evaluation strategy to determine the final data item commonality metric based on the characteristics and importance of the three metrics. This strategy may include weighted average, linear combination or other complex algorithms. For example, the computer system can assign a weight to each metric, which can be determined based on domain knowledge, data characteristics and the needs of the prediction model. Then, the computer system weighted sums the three metrics according to these weights to obtain the final data item commonality metric.

[0243] For example, suppose the commonality metric value of the first data item is 0.8, the commonality metric value of the second data item is 0.7, and the commonality metric value of the third data item is 0.6. The computer system can assign weights of 0.4, 0.3, and 0.3 to these three metrics, respectively, based on domain knowledge and data characteristics. Then, the final commonality metric value of the data item is obtained by weighted summation calculation: 0.8*0.4+0.7*0.3+0.6*0.3=0.71. This value comprehensively reflects the overall similarity or difference between the samples of the group without hyperamylasemia and the samples of the group with hyperamylasemia in the time series data. Through the processing of step S20433, the computer system can obtain a quantitative commonality metric value of the data item, which not only takes into account the information of the overlapping data items, but also incorporates the coverage and completeness of the time series data and the importance of the risk monitoring indicators, providing a comprehensive and accurate reference basis for subsequent model training and data analysis. This will help improve the accuracy and generalization ability of the prediction model and provide doctors with more accurate postoperative risk assessment.

[0244] As an implementation manner, step S600, training an initial post-ERCP hyperamylasemia prediction model based on a modified post-ERCP monitoring data sample library to obtain a target post-ERCP hyperamylasemia prediction model, may specifically include:

[0245] Step S610: inputting each modified post-ERCP monitoring data sample in the modified post-ERCP monitoring data sample library into the initial post-ERCP hyperamylasemia prediction model to obtain a monitoring data representation vector corresponding to each modified post-ERCP monitoring data sample;

[0246] Step S620: determining the hyperamylasemia prediction result corresponding to each modified ERCP post-operative monitoring data sample according to the monitoring data characterization vector corresponding to each modified ERCP post-operative monitoring data sample;

[0247] Step S630: Based on the hyperamylasemia prediction results corresponding to each revised ERCP post-operative monitoring data sample and the errors between the hyperamylasemia markers corresponding to each revised ERCP post-operative monitoring data sample, the model parameters of the initial post-ERCP hyperamylasemia prediction model are optimized to obtain the target post-ERCP hyperamylasemia prediction model.

[0248] In the embodiment of the present application, the goal of step S600 is to train an initial prediction model through the modified ERCP post-operative monitoring data sample library, and obtain the final target prediction model through iterative optimization.

[0249] Step S610 involves inputting a corrected data sample and obtaining a monitoring data characterization vector. In this step, the computer system inputs each corrected data sample in the corrected post-ERCP monitoring data sample library into the initial post-ERCP hyperamylasemia prediction model. This model is usually a pre-designed machine learning model, such as logistic regression, random forest or deep learning network, which can output corresponding prediction results based on the input monitoring data. When the data sample is input into the model, the model converts it into a series of monitoring data characterization vectors. These vectors are the internal representation of the input data in the model, which capture the key features in the data and are used for subsequent prediction calculations. For example, if a data sample contains two monitoring indicators, serum amylase level and abdominal pain score, the model can convert these two indicators into a two-dimensional characterization vector, such as [150,2], where 150 represents the serum amylase level and 2 represents the abdominal pain score.

[0250] In step S620, the prediction result is determined based on the characterization vector. After obtaining the monitoring data characterization vector corresponding to each corrected data sample, the computer system uses these vectors to calculate the hyperamylasemia prediction result corresponding to each sample. This process usually involves the calculation logic and parameter settings within the model. For different machine learning models, this calculation process may be different. Taking the logistic regression model as an example, it can use the characterization vector and the weight parameters of the model to calculate a prediction probability, which represents the possibility that the sample belongs to the hyperamylasemia category. If the probability exceeds a preset threshold (such as 0.5), the model will predict the sample as hyperamylasemia occurs; otherwise, it is predicted as not occurring.

[0251] In step S630, the model parameters are optimized based on the prediction error. After obtaining the prediction results of each corrected data sample, the computer system compares these prediction results with the true hyperamylasemia markers (i.e., true labels) corresponding to the samples, and calculates the error between them. This error reflects the accuracy of the model prediction and is a key indicator for model optimization. Next, the computer system uses this error to optimize the parameters of the initial prediction model (such as weight parameters, bias terms, etc.). This process usually involves optimization algorithms such as gradient descent, which minimize the prediction error by iteratively adjusting the model parameters. As the iteration proceeds, the prediction accuracy of the model will gradually increase until the preset stop condition is reached (such as the error is less than a certain threshold, the number of iterations reaches an upper limit, etc.). Finally, after multiple iterative optimizations, the computer system obtains a target ERCP postoperative hyperamylasemia prediction model. This model has a high prediction accuracy and can accurately predict the risk of hyperamylasemia in patients after ERCP, providing valuable reference information for doctors.

[0252] For example, assume that the initial prediction model is a simple logistic regression model that uses serum amylase levels and abdominal pain scores as input features. In step S610, the computer system inputs the corrected data samples into the model to obtain a characterization vector for each sample. Then, in step S620, the model uses these characterization vectors to calculate the prediction probability of each sample and converts it into a prediction result based on a threshold. Next, in step S630, the computer system calculates the error between the prediction result and the true label, and optimizes the weight parameters of the model using a gradient descent algorithm. After multiple iterations, the prediction accuracy of the model was significantly improved, and a target prediction model was finally obtained. This model can accurately predict the risk of hyperamylasemia in patients after ERCP, providing strong support for doctors.

[0253] The present application embodiment also provides an application process of a target post-ERCP hyperamylasemia prediction model, which may specifically include the following steps:

[0254] Step S1: obtaining the post-ERCP monitoring data of the cases to be analyzed;

[0255] Step S2: inputting the post-ERCP monitoring data to be analyzed into the target post-ERCP hyperamylasemia prediction model to obtain the prediction result of the occurrence of hyperamylasemia;

[0256] Step S3: If the prediction result of the occurrence of hyperamylasemia indicates that the post-ERCP monitoring data to be analyzed is the occurrence of hyperamylasemia, the case to be analyzed is determined as a high-risk case.

[0257] In the postoperative management of ERCP (endoscopic retrograde pancreatocholangiopancreatography), the prediction of hyperamylasemia is a key task because it can help doctors identify patients who may be at higher risk in advance, thereby formulating more personalized treatment plans. Based on the training process of the postoperative hyperamylasemia prediction model mentioned above, the embodiment of the present application also provides the application process of the model.

[0258] In step S1, the post-ERCP monitoring data of the case to be analyzed is obtained. In this step, the computer system first needs to obtain the post-ERCP monitoring data of the case to be analyzed. These data usually include various monitoring indicators of the patient at different time points after surgery, such as serum amylase level, abdominal pain score, operation time, etc. These data reflect the patient's postoperative recovery and possible risk factors. For example, a doctor can perform a series of postoperative monitoring on a patient who has just completed an ERCP operation, and input these monitoring data into the computer system for analysis.

[0259] In step S2, the post-ERCP monitoring data to be analyzed is input into the target post-ERCP hyperamylasemia prediction model to obtain the prediction result of the occurrence of hyperamylasemia. After obtaining the post-ERCP monitoring data of the case to be analyzed, the computer system inputs these data into the previously trained target post-ERCP hyperamylasemia prediction model. This model has been trained and optimized through a large amount of historical data, and can accurately predict the risk of hyperamylasemia in patients based on the input monitoring data. The model will perform complex calculations and analyses on the input data, and finally output a prediction result of the occurrence of hyperamylasemia. This result is usually a probability value, indicating the possibility of hyperamylasemia in the patient.

[0260] In step S3, if the prediction result of the occurrence of hyperamylasemia indicates that the post-ERCP monitoring data to be analyzed is the occurrence of hyperamylasemia, the case to be analyzed is determined as a high-risk case. After obtaining the prediction result of the occurrence of hyperamylasemia, the computer system compares it with a preset threshold. This threshold is usually set according to clinical needs and model performance, and is used to determine whether the patient is a high-risk case. If the prediction result exceeds this threshold (for example, the prediction probability is greater than 0.5), the computer system will consider that the patient has a high risk of hyperamylasemia and will determine it as a high-risk case. In this way, doctors can take closer observations and targeted treatment measures for high-risk patients based on this prediction result, thereby reducing the incidence of postoperative complications and improving the patient's treatment effect and quality of life. Through the above steps, the application process of the target ERCP post-operative hyperamylasemia prediction model provided in the embodiment of the present application can help doctors more accurately assess the patient's postoperative risks and provide strong support for clinical decision-making.

[0261] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

Claims

1. A method for training a prediction model for hyperamylasemia after ERCP, characterized in that: The method comprises: Obtaining a sample library of post-ERCP monitoring data corresponding to the training case library; the sample library of post-ERCP monitoring data includes samples of a group in which hyperamylasemia does not occur and samples of a group in which hyperamylasemia occurs; the samples of the group in which hyperamylasemia does not occur carry a marker of hyperamylasemia not occurring, and the samples of the group in which hyperamylasemia occurs carry a marker of hyperamylasemia occurring; Determine a monitoring indicator representation vector corresponding to each ERCP post-operative monitoring data sample in the ERCP post-operative monitoring data sample library; Grouping the plurality of monitoring indicator representation vectors corresponding to the ERCP post-operative monitoring data sample library to obtain a plurality of groups; Determining a target group based on the monitoring indicator representation vectors corresponding to the multiple groups respectively; Correcting the markers of the samples in the group with no hyperamylasemia corresponding to the target group to the markers of hyperamylasemia occurrence, and obtaining a corrected ERCP postoperative monitoring data sample library; Training an initial post-ERCP hyperamylasemia prediction model based on the modified post-ERCP monitoring data sample library to obtain a target post-ERCP hyperamylasemia prediction model; Wherein, the plurality of monitoring indicator representation vectors corresponding to the ERCP postoperative monitoring data sample library are grouped to obtain a plurality of groups, including: Determine the commonality measurement value of the monitoring indicators of the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group according to the monitoring indicator characterization vectors corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group respectively; Grouping the common measurement values ​​of multiple monitoring indicators corresponding to the ERCP post-operative monitoring data sample library to obtain the multiple groups; The determining the target group based on the monitoring indicator characterization vectors respectively corresponding to the multiple groups includes: Determine a target group based on the common measurement values ​​of the monitoring indicators corresponding to the multiple groups respectively; The determining the target group based on the monitoring indicator characterization vectors respectively corresponding to the multiple groups includes: Determining a group including the danger monitoring indicator characterization vector among the multiple groups as a target group; The hazard monitoring indicator characterization vector is determined by the following process: Obtaining a preset ERCP post-operative monitoring database corresponding to a preset case database; the preset ERCP post-operative monitoring database includes monitoring data before and after the occurrence of symptoms; Determine a monitoring indicator representation vector corresponding to each preset ERCP postoperative monitoring data in the preset ERCP postoperative monitoring database; Grouping multiple preset monitoring indicator representation vectors corresponding to the preset ERCP postoperative monitoring database to obtain multiple preset groups; Determine the percentage of symptom occurrence monitoring data corresponding to each preset group; Determine the preset group whose percentage of symptom occurrence monitoring data is greater than the percentage threshold as a risk group; Determining a preset monitoring indicator characterization vector in the risk group as the risk monitoring indicator characterization vector; The method of grouping the plurality of preset monitoring indicator representation vectors corresponding to the preset ERCP postoperative monitoring database to obtain a plurality of preset groups includes: Determine a commonality measurement value of the preset monitoring indicators of the monitoring data before the symptom occurs and the monitoring data after the symptom occurs, according to the preset monitoring indicator characterization vectors respectively corresponding to the monitoring data before the symptom occurs and the monitoring data after the symptom occurs; Grouping based on the common measurement values ​​of multiple preset monitoring indicators corresponding to the preset ERCP postoperative monitoring database to obtain the multiple preset groups; The method further comprises: Determine the commonality metric value of the selected monitoring indicator in the risk group as the commonality metric threshold; the commonality metric value of the selected monitoring indicator in the risk group is less than the first commonality metric value; the first commonality metric value is the commonality metric value in the risk group except the commonality metric value of the selected monitoring indicator; The determining of the target group based on the commonality measurement values ​​of the monitoring indicators respectively corresponding to the plurality of groups includes: Determine a common metric value of the target monitoring indicator in each group; the common metric value of the target monitoring indicator in any group is less than a second common metric value in any group; the second common metric value in any group is a common metric value in any group except the common metric value of the target monitoring indicator; Determine a group whose target monitoring indicator commonality metric value is not less than the commonality metric threshold as the target group; The step of determining the monitoring indicator representation vector corresponding to each ERCP post-operative monitoring data sample in the ERCP post-operative monitoring data sample library includes: Retrieving a target representation vector mining model, performing representation vector mining on each ERCP post-operative monitoring data sample in the ERCP post-operative monitoring data sample library, and obtaining the monitoring indicator representation vector; The target representation vector mining model is trained by the following steps: Acquire monitoring indicator data corresponding to a plurality of post-ERCP monitoring data samples, wherein the post-ERCP monitoring data samples include a plurality of monitoring indicator data; Based on the representation vector mining model, the monitoring indicator data corresponding to the multiple post-ERCP monitoring data samples are respectively subjected to representation vector mining, and the monitoring indicator representation vectors obtained by mining are projected into the representation vector domain to obtain the vector positions of the monitoring indicator representation vectors corresponding to the multiple post-ERCP monitoring data samples in the representation vector domain; According to the vector positions of the monitoring indicator representation vectors corresponding to the multiple ERCP post-operative monitoring data samples in the representation vector domain, cluster analysis is performed on the monitoring indicator representation vectors corresponding to the multiple ERCP post-operative monitoring data samples to obtain multiple classification clusters, wherein the classification clusters include multiple monitoring indicator representation vectors that satisfy vector space similarity; Performing monitoring indicator occurrence rate analysis on monitoring indicator data corresponding to monitoring indicator representation vectors in the same classification cluster to obtain cluster core indicators corresponding to the classification cluster; According to the cluster core indicators and the monitoring indicator feature mining focus of the characterization vector mining model, the characterization vector mining model is trained until the characterization vector mining model reaches a preset convergence condition to obtain a trained characterization vector mining model; Determining the data item commonality measurement value of the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group according to the data item sequences respectively corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group includes: Determine a target overlapping data sequence of the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group according to the data item sequences respectively corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group; Determine the length of the data item sequence corresponding to the samples of the hyperamylasemia non-occurring group and the samples of the hyperamylasemia occurring group respectively; Determine a first data item commonality metric value between the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group based on the lengths of the data item sequences respectively corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group and the target overlapping data sequence; Based on the first data item commonality metric value, a data item commonality metric value between the hyperamylasemia non-occurring group samples and the hyperamylasemia occurring group samples is determined.

2. The method according to claim 1, characterized in that The method further comprises: Determine the data item sequences corresponding to the samples in the hyperamylasemia non-occurring group and the samples in the hyperamylasemia occurring group respectively; Determine a data item commonality measurement value between the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group according to the data item sequences respectively corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group; The grouping of the common measurement values ​​of the multiple monitoring indicators corresponding to the ERCP post-operative monitoring data sample library to obtain the multiple groups includes: The plurality of monitoring indicator commonality measurement values ​​and the plurality of data item commonality measurement values ​​corresponding to the ERCP postoperative monitoring data sample library are grouped to obtain the plurality of groups.

3. The method according to claim 1, characterized in that The monitoring indicator occurrence rate analysis is performed on the monitoring indicator data corresponding to the monitoring indicator representation vector in the same classification cluster to obtain the cluster core indicator corresponding to the classification cluster, including: Performing monitoring indicator occurrence rate analysis on the monitoring indicator data corresponding to the monitoring indicator representation vectors in the same classification cluster to obtain the occurrence rate of each of the monitoring indicator data; When the appearance rate of the target monitoring indicator data reaches a preset appearance rate, the target monitoring indicator data is determined as the cluster core indicator corresponding to the classification cluster.

4. The method according to claim 1 or 3, characterized in that: The training of the representation vector mining model based on the cluster core indicators and the monitoring indicator feature mining focus of the representation vector mining model includes: When the cluster core indicator does not correspond to the monitoring indicator feature mining focus corresponding to the representation vector mining model, the error function for training the representation vector mining model is optimized according to the monitoring indicator feature mining focus; and the representation vector mining model is trained based on the optimized error function; or; When the cluster core indicator does not correspond to the monitoring indicator feature mining focus corresponding to the representation vector mining model, a post-ERCP monitoring data sample set is obtained according to the monitoring indicator feature mining focus, and the post-ERCP monitoring data samples in the post-ERCP monitoring data sample set are samples corresponding to the monitoring indicator feature mining focus; the representation vector mining model is trained based on the post-ERCP monitoring data samples in the post-ERCP monitoring data sample set; or; When the cluster core indicator does not correspond to the monitoring indicator feature mining focus corresponding to the representation vector mining model, when obtaining the monitoring indicator data corresponding to multiple ERCP post-operative monitoring data samples, the monitoring indicator data corresponding to the monitoring indicator feature mining focus is added to the monitoring indicator data.

5. The method according to claim 4, characterized in that According to the vector positions of the monitoring indicator representation vectors corresponding to the multiple ERCP post-operative monitoring data samples in the representation vector domain, cluster analysis is performed on the monitoring indicator representation vectors corresponding to the multiple ERCP post-operative monitoring data samples to obtain multiple classification clusters, including: Determining the spatial similarity between the monitoring indicator representation vectors according to the distribution positions of the monitoring indicator representation vectors corresponding to the plurality of post-ERCP monitoring data samples in the representation vector domain; When the spatial similarity is greater than a spatial similarity threshold, it is determined that the monitoring indicator representation vectors whose spatial similarity is less than the spatial similarity threshold are located in the same classification cluster.

6. The method according to claim 1, characterized in that The determining, based on the first data item commonality metric value, of the data item commonality metric value between the samples in the hyperamylasemia non-occurring group and the samples in the hyperamylasemia occurring group comprises: Determine the overlapping data items of the samples of the non-hyperamylasemia group and the samples of the hyperamylasemia group according to the data item sequences respectively corresponding to the samples of the non-hyperamylasemia group and the samples of the hyperamylasemia group; Determine a second data item commonality metric value between the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group based on the lengths of the data item sequences corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group and the overlapping data items; Based on the first data item commonality metric value and the second data item commonality metric value, the data item commonality metric value of the hyperamylasemia non-occurring group samples and the hyperamylasemia occurring group samples is determined.

7. The method according to claim 6, characterized in that The determining, based on the first data item commonality metric value and the second data item commonality metric value, of the data item commonality metric value between the samples in the hyperamylasemia non-occurring group and the samples in the hyperamylasemia occurring group comprises: Determine the percentages of risk monitoring indicators corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group, respectively, according to the data item sequences corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group; Determine the third data item commonality measurement value of the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group according to the percentages of the risk monitoring indicators respectively corresponding to the samples in the non-hyperamylasemia group and the samples in the hyperamylasemia group; Based on the first data item commonality metric value, the second data item commonality metric value, and the third data item commonality metric value, the data item commonality metric value of the non-hyperamylasemia group samples and the hyperamylasemia group samples is determined.

8. The method according to claim 1, characterized in that The method of training an initial post-ERCP hyperamylasemia prediction model based on the modified post-ERCP monitoring data sample library to obtain a target post-ERCP hyperamylasemia prediction model comprises: Inputting each modified post-ERCP monitoring data sample in the modified post-ERCP monitoring data sample library into the initial post-ERCP hyperamylasemia prediction model to obtain a monitoring data representation vector corresponding to each modified post-ERCP monitoring data sample; Determine the hyperamylasemia prediction result corresponding to each of the modified ERCP post-operative monitoring data samples according to the monitoring data characterization vector corresponding to each of the modified ERCP post-operative monitoring data samples; Based on the hyperamylasemia prediction results corresponding to each of the modified post-ERCP monitoring data samples and the errors between the hyperamylasemia markers corresponding to each of the modified post-ERCP monitoring data samples, the model parameter variables of the initial post-ERCP hyperamylasemia prediction model are optimized to obtain the target post-ERCP hyperamylasemia prediction model; The method further comprises: Obtain the post-ERCP monitoring data of the cases to be analyzed; Inputting the post-ERCP monitoring data to be analyzed into the target post-ERCP hyperamylasemia prediction model to obtain a hyperamylasemia occurrence prediction result; If the prediction result of the occurrence of hyperamylasemia indicates that the post-ERCP monitoring data to be analyzed is the occurrence of hyperamylasemia, the case to be analyzed is determined as a high-risk case.

9. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.