Prediction system for common bile duct stones recurrence after ERCP based on machine learning

Abstract

The present invention discloses a machine learning-based system for predicting the recurrence of common bile duct stones after ERCP. The system comprises the following steps: a data acquisition module for forming an original clinical data set; a data preprocessing module for constructing a standardized preprocessed clinical data set; a feature construction and hyperparameter search space definition module for defining the hyperparameter search space of an improved bidirectional gated recurrent unit network model; an improved bidirectional gated recurrent unit network model construction module for outputting a recurrence risk prediction value; a Gray Wolf optimization hyperparameter tuning module for ultimately generating an optimized improved bidirectional gated recurrent unit network model; a model training and prediction module for constructing risk classification rules based on predicted probabilities and postoperative key indicators; and an individualized result output module for outputting and delivering the predicted results to the physician. The present invention has significant clinical value in the early identification of high-risk individuals in actual deployment.

Description

Technical Field

[0001] The present invention relates to the technical field of common bile duct stones, and in particular to a machine learning-based common bile duct stone recurrence prediction system after ERCP. Background Art

[0002] With the development of minimally invasive technology, endoscopic retrograde cholangiopancreatography has become an important means of treating common bile duct stones. However, the problem of postoperative recurrence still plagues clinical practice. Some patients develop stones again within months to years after surgery, which places a heavy burden on patients' health and the medical system. In order to effectively predict the risk of recurrence and optimize follow-up strategies, clinical research has gradually introduced data-driven methods in recent years to attempt risk modeling and predictive analysis.

[0003] Current predictions of postoperative recurrence of common bile duct stones mainly rely on retrospective statistical analysis or risk assessment models based on expert experience. These models use single or a small number of clinical characteristics as risk factors, such as patient age, gender, postoperative biliary dilatation, or liver function indicators. These models make it difficult to systematically integrate multidimensional data from patients before, during, and after surgery. On the other hand, some studies have attempted to use traditional machine learning methods such as logistic regression and support vector machines for prediction. However, their modeling capabilities are limited by insufficient feature expression capabilities and non-temporal data structures, making them unable to accurately capture dynamic risk signals during intraoperative status changes and postoperative recovery.

[0004] In addition, existing prediction methods generally lack an optimization mechanism for model parameters and often rely on manual experience to set parameters, making it difficult to achieve the optimal configuration of model performance. There is still much room for improvement in prediction accuracy, individual adaptability, and time series modeling capabilities. Especially in postoperative risk stratification and individualized management, existing technologies are difficult to meet the requirements of precision medicine. Therefore, there is an urgent need for a recurrence prediction method that can integrate multi-stage clinical data, has time series modeling capabilities, and supports adaptive optimization of model parameters to improve prediction accuracy and clinical practicality. Summary of the Invention

[0005] One purpose of the present invention is to propose a machine learning-based system for predicting the recurrence of common bile duct stones after ERCP. The present invention has significant clinical value in the early identification of high-risk individuals in actual deployment.

[0006] According to an embodiment of the present invention, a machine learning-based system for predicting recurrence of common bile duct stones after ERCP includes the following modules:

[0007] The data acquisition module is used to collect multidimensional clinical data of patients before, during and after ERCP to form the original clinical data set;

[0008] A data preprocessing module is used to perform preprocessing operations on the original clinical data set and construct a standardized preprocessed clinical data set;

[0009] Feature construction and hyperparameter search space definition module, which is used to construct preoperative, intraoperative, and postoperative stage feature vectors based on standardized preprocessed clinical datasets and define the hyperparameter search space of the improved bidirectional gated recurrent unit network model;

[0010] Improved bidirectional gated recurrent unit network model construction module, used to construct an improved bidirectional gated recurrent unit network model that integrates a temporal enhancement mechanism, a clinical stage-aware gating mechanism, and a risk-guided fusion strategy to output a recurrence risk prediction value;

[0011] The Gray Wolf Optimization hyperparameter tuning module is used to initialize the hyperparameter search space based on key medical features, construct a medically guided initialization population, and perform multiple rounds of Gray Wolf Optimization searches based on a stage-by-stage hierarchical convergence strategy and a risk-sensitive fitness function to obtain the optimal hyperparameter configuration, ultimately generating an optimized improved bidirectional gated recurrent unit network model.

[0012] A model training and prediction module is used to use the optimized improved bidirectional gated recurrent unit network model to perform forward prediction on standardized new patient clinical data, output a recurrence probability value, and construct a risk classification rule based on the predicted probability and postoperative key indicators;

[0013] The individualized result output module is used to stratify the prediction results into high-risk, medium-risk, and low-risk levels. The output results include: individual recurrence probability value, stratified risk level, and explanation of key driving medical indicators. The prediction results are also pushed to the doctor-side clinical decision support system to assist in formulating personalized postoperative management strategies and follow-up plans.

[0014] A method for predicting the recurrence of common bile duct stones after ERCP based on machine learning is applied to a system for predicting the recurrence of common bile duct stones after ERCP based on machine learning, comprising the following steps:

[0015] S1. Collect multidimensional clinical data sets from patients before, during, and after ERCP to form a clinical data set for prediction.

[0016] S2. Preprocess the clinical dataset and construct a standardized preprocessed clinical dataset;

[0017] S3. Construct feature vectors based on the preprocessed clinical dataset to establish a feature dataset suitable for the input requirements of a bidirectional gated recurrent unit network.

[0018] S4. Build an improved bidirectional gated recurrent unit network model to predict recurrence risk;

[0019] S5. Based on key indicators reflecting the recurrence of common bile duct stones after ERCP in the preprocessed clinical dataset, a medically guided initialization of the hyperparameter search space was performed to form an optimized initial population. A global search for the hyperparameters of the improved bidirectional gated recurrent unit network model was performed using the Grey Wolf Optimization Algorithm. Dynamic fitness evaluation was performed, and the population position was updated to obtain the optimal hyperparameters. This optimal hyperparameter configuration was applied to the improved bidirectional gated recurrent unit network model to obtain the optimized improved bidirectional gated recurrent unit network model.

[0020] S6. Use the optimized improved bidirectional gated recurrent unit network model to predict the recurrence risk of new preprocessed clinical datasets and output individualized prediction results for the recurrence risk of common bile duct stones after ERCP.

[0021] Optionally, the S1 includes the following steps:

[0022] S11. Set the clinical data collection window to collect the patient's multi-dimensional clinical raw data before, during and after ERCP to form the initial clinical data set :

[0023] ;

[0024] in, Indicates the Clinical data records of patients, is the total number of patients, Indicates the The surgery time of the patients Represents the preoperative data vector, which includes patient demographic information and basic examination indicators. Represents the intraoperative data vector, which includes variables related to surgical records. Represents the postoperative data vector, which includes biochemical indicators, imaging data and follow-up record characteristics;

[0025] S12. For initial clinical data collection A unified feature space mapping function is defined for different types of clinical data. The feature space mapping function includes unified data encoding, time alignment, and structured rearrangement operations. The feature space mapping function maps each type of data to a standardized feature space to form preoperative feature sets, intraoperative feature sets, and postoperative feature sets.

[0026] S13. Integrate the preoperative feature set, intraoperative feature set, and postoperative feature set in chronological order to construct a clinical dataset for modeling. Normalize the timestamp set in the clinical dataset to form a standardized time axis for calibrating the time series alignment operation between patients. and timeline as standardized input.

[0027] Optionally, the S2 includes the following steps:

[0028] S21. Clinical Datasets The eigenvectors in Perform missing value detection and processing, and use the mean filling strategy to fill in the missing value feature dimensions. The patient's When a feature value is missing, the value of the first The arithmetic mean of the eigenvalues is replaced to form the eigenvector after the missing values are filled;

[0029] S22. Based on the missing value filled feature vector, perform outlier detection and processing. For each feature dimension, calculate the median of all patients' feature values and the absolute value of the deviation from the median. Use this to identify outliers. When the deviation between a feature value and the median exceeds a preset tolerance multiple, replace the feature value with the median of the current feature dimension, forming a feature vector after outlier processing.

[0030] S23. Perform normalization on the feature vector after outlier processing, and standardize all feature values to the interval Within the range, the minimum value of the feature dimension in all patients is used as the lower limit and the maximum value is used as the upper limit. Linear scaling is performed on each feature value to form a normalized feature vector.

[0031] S24. Express all normalized patient feature vectors as the standardized preprocessed clinical feature matrix:

[0032] ;

[0033] in, Indicates the Patient No. The normalized value of the feature, The dimension is The standardized pretreatment clinical characteristics matrix, is the number of patient samples, is the feature dimension;

[0034] S25. Combine the standardized preprocessed clinical feature matrix with the standardized time axis to construct a time-aligned standardized preprocessed clinical dataset .

[0035] Optionally, S3 includes the following steps:

[0036] S31. Based on standardized preprocessed clinical dataset , extract the clinical time series feature vector of each patient , combine the feature vectors of all patients in chronological order to construct a two-dimensional input feature dataset, which has OK, The structure of the columns, represents the number of patients, represents the standardized characteristic dimensions of each patient;

[0037] S32. Convert the two-dimensional input feature data set into a sequence format divided by time steps, and set the time step length to , divide the feature vector of each patient into continuous time step input, each time step input contains a normalized feature dimension consistent eigenvectors;

[0038] S33. Establishing a Hyperparameter Search Space for an Improved Bidirectional Gated Recurrent Unit Network Model , the hyperparameter search space includes the number of network layers , the number of hidden units in each layer , learning rate , dropout rate and batch processing scale ,in:

[0039] The number of network layers is a positive integer, which is used to define the depth of the GRU layer in the model;

[0040] The number of hidden units in each layer is a positive integer, which is used to define the number of neurons in each layer of GRU units;

[0041] The learning rate is a continuous real number between 0 and 1, which is used to control the step size of model parameter updates;

[0042] The dropout rate is a continuous real number between 0 and 1, which is used to control the random dropout ratio of nodes in the model;

[0043] The batch size is a positive integer that sets the number of samples used for each round of parameter update in model training.

[0044] S34. For each hyperparameter Set the corresponding initial search range, the range of the number of network layers is set from the minimum number of layers to the maximum number of layers, the range of the number of hidden units in each layer is set from the minimum number of hidden units to the maximum number of hidden units, the range of the learning rate is set from the minimum learning rate to the maximum learning rate, the range of the dropout rate is set from the minimum dropout rate to the maximum dropout rate, and the range of the batch size is set from the minimum batch size to the maximum batch size.

[0045] Optionally, the S4 includes the following steps:

[0046] S41. Based on preoperative data vector , intraoperative data vector and postoperative data vector , set clinical stage labels for each time segment ,in Indicates the corresponding stage;

[0047] S42. Feature vector for each stage , calculate the changing trend of clinical characteristics between this stage and the previous stage, and form a time series enhanced feature vector ;

[0048] S43. Enhance the time series feature vector with clinical stage labels Input into clinical stage perception gating mechanism and define gating weights for:

[0049] ;

[0050] in, is the Sigmoid function, 、 、 are the weights, stage embedding matrices, and biases learned during model training, The gating weight for each feature dimension represents the importance of this indicator to recurrence risk modeling at this stage;

[0051] S44. Construct a bidirectional gated recurrent unit network with the input being the feature vector after stage regulation :

[0052] ;

[0053] in, It is the Hadamard element-wise product, indicating that only the feature dimensions that are highly correlated with the risk of recurrence are retained in a specific stage after ERCP.

[0054] S45. The characteristic vector after stage regulation As input, the In the network structure composed of layers of bidirectional gated recurrent units, each layer is set hidden units, outputting the forward state and backward state , combined with the temporal significance of post-ERCP recurrence in the postoperative period, a risk-guided bidirectional fusion mechanism is used to define the fusion representation :

[0055] ;

[0056] in, is the risk attention weight;

[0057] S46. Fusion representation of all time step outputs Perform weighted pooling to obtain the patient-level global temporal feature expression vector ;

[0058] S47. Express the global temporal feature vector Input to the fully connected layer and predict the output recurrence risk :

[0059] ;

[0060] in, 、 are weight and bias terms, Indicates the The recurrence risk of each patient;

[0061] S48. During the training process, hyperparameters are used for optimization control:

[0062] Learning rate : Control parameter update speed;

[0063] Dropout rate : Used to randomly drop some neurons in the bidirectional gated recurrent unit layer and the fully connected layer;

[0064] Batch size : Control the number of samples selected in each round of parameter update;

[0065] Number of network layers , the number of hidden units : Controlling the structural depth and feature capacity of bidirectional gated recurrent unit networks.

[0066] Optionally, the S5 includes the following steps:

[0067] S51. Extracting a set of key medical features significantly associated with the risk of recurrence after common bile duct stone surgery based on standardized preprocessed clinical datasets The key medical feature set includes the total number of stones removed during surgery, the change rate of TBiL after surgery, the peak value of ALT after surgery, and the results of the 7-day follow-up after surgery. The feature strength vector is constructed ;

[0068] S52. Based on feature intensity vector , introduce the medical feature-guided initialization mechanism to generate the optimized population initialization offset matrix , and inject the optimized population initialization offset matrix into the generation process of the initial population of the gray wolf algorithm:

[0069] ;

[0070] in, is the bias intensity weight, Initialize the hyperparameters for the standard uniform individual and optimize the population initialization offset matrix Depend on Mapping generation allows the hyperparameters corresponding to high-risk medical features to enter the search space first, Optimizing the population size in the algorithm for gray wolves;

[0071] S53. Partition the search space by clinical stage, construct a stage-by-stage convergence strategy, and define the search state label for each individual gray wolf in the population. , dynamically control the update step factor according to the search status label , given the high-risk postoperative period, prioritize maintaining global search capabilities:

[0072] ;

[0073] in, is the maximum number of iterations, is the initial step size control factor, t represents the current iteration index in the gray wolf optimization algorithm;

[0074] S54. Based on the training set and the validation set, for each set of candidate hyperparameters Constructing a risk-sensitive dynamic fitness function:

[0075]

[0076] in, Represents the area under the ROC curve of the model on the validation set, represents the recurrence sensitivity of the model on samples in the postoperative period, is the balance coefficient;

[0077] S55. According to the gray wolf optimization strategy, select the best individual, the second best individual, and the second best individual from the current population, and set their positions respectively. 、 、 , then update the The position of each individual is:

[0078]

[0079] in, is the combined Euclidean distance between the current individual and the three best individuals;

[0080] S56. Repeat steps S53-S55, iteratively perform global search until the maximum number of iterations is reached Or the fitness change is less than the preset threshold The convergence conditions of , and finally the optimal hyperparameter configuration is obtained:

[0081]

[0082] S57. Configure the optimal hyperparameters Applied to the bidirectional gated recurrent unit network structure, the final training task is globally optimized, the optimal hyperparameter configuration is applied to the improved bidirectional gated recurrent unit network model, and the preprocessed clinical dataset is iteratively trained until the training of the improved bidirectional gated recurrent unit network model reaches the preset stopping criterion, and the optimized improved bidirectional gated recurrent unit network model is obtained.

[0083] Optionally, the S6 includes the following steps:

[0084] S71. Input the preprocessed new clinical data set into the optimized improved bidirectional gated recurrent unit network model, perform forward prediction, and obtain the first Output value of stone recurrence probability for new patients ;

[0085] S72. Output value of stone recurrence probability for new patients , combined with the high-risk index threshold after surgery, to construct an individualized recurrence risk classification standard based on medical rules, and divide each patient into the corresponding risk level;

[0086] S73. Output the probability of stone recurrence for each new patient The key indicator data during and after the operation are correlated and statistically analyzed, and the data are stratified by risk level to form an individualized prediction result of the risk of recurrence of common bile duct stones after ERCP.

[0087] Optionally, the risk level includes:

[0088] High-risk patients are classified as high risk if they meet any of the following conditions: Predicted probability , and the TBiL change rate within 7 days after surgery , or postoperative ALT peak , or the number of stones remaining during surgery ;

[0089] Patients at intermediate risk are classified as those who meet any of the following conditions: Predicted probability , and the TBiL change rate within 7 days after surgery , or the number of stones remaining during surgery , or postoperative ALT peak ;

[0090] Low-risk patients are classified as low-risk if they meet any of the following conditions: Predicted probability , and the TBiL change rate after surgery , and the peak ALT level after surgery , and there are no residual stones during the operation, that is .

[0091] The beneficial effects of the present invention are:

[0092] (1) The present invention constructs data input channels for the preoperative, intraoperative, and postoperative stages, and combines the changing trend of the feature vectors in each stage to generate time series enhancement features. On this basis, a clinical stage perception gating mechanism is proposed. By introducing label embedding and feature gating weight learning for each stage, the expression intensity of each feature dimension in different time segments is dynamically adjusted, effectively emphasizing the contribution of postoperative biochemical indicators and imaging information to recurrence modeling, and the feature flow direction can be automatically adjusted according to the stage risk weight, effectively improving the model's ability to pay attention to recurrence-sensitive areas.

[0093] (2) The present invention embeds key medical features into the initial population construction of the gray wolf optimization algorithm to form an initialization offset matrix driven by medical features, and introduces a stage-by-stage convergence control strategy to maintain global search capabilities in the early stage of the high-risk postoperative stage and quickly converge in the preoperative stage to stabilize the basic network.

[0094] (3) The present invention constructs a risk-guided bidirectional fusion mechanism based on GRU, introduces key stage weights in the network output stage, forms a temporal aggregation representation of dynamic risk perception, and combines weighted pooling to enhance the model's ability to integrate long-term dependent information. In actual deployment, it has significant clinical value for the early identification of high-risk individuals. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0096] Figure 1 This is a flowchart of a machine learning-based system for predicting the recurrence of common bile duct stones after ERCP proposed by the present invention. DETAILED DESCRIPTION

[0097] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0098] refer to Figure 1 , a machine learning-based prediction system for common bile duct stone recurrence after ERCP, including the following modules:

[0099] The data acquisition module is used to collect multidimensional clinical data of patients before, during and after ERCP to form the original clinical data set;

[0100] A data preprocessing module is used to perform preprocessing operations on the original clinical data set and construct a standardized preprocessed clinical data set;

[0101] Feature construction and hyperparameter search space definition module, which is used to construct preoperative, intraoperative, and postoperative stage feature vectors based on standardized preprocessed clinical datasets and define the hyperparameter search space of the improved bidirectional gated recurrent unit network model;

[0102] Improved bidirectional gated recurrent unit network model construction module, used to construct an improved bidirectional gated recurrent unit network model that integrates a temporal enhancement mechanism, a clinical stage-aware gating mechanism, and a risk-guided fusion strategy to output a recurrence risk prediction value;

[0103] The Gray Wolf Optimization hyperparameter tuning module is used to initialize the hyperparameter search space based on key medical features, construct a medically guided initialization population, and perform multiple rounds of Gray Wolf Optimization searches based on a stage-by-stage hierarchical convergence strategy and a risk-sensitive fitness function to obtain the optimal hyperparameter configuration, ultimately generating an optimized improved bidirectional gated recurrent unit network model.

[0104] The model training and prediction module is used to use the optimized improved bidirectional gated recurrent unit network model to perform forward predictions on standardized new patient clinical data, output recurrence probability values, and construct risk classification rules based on the predicted probability and key postoperative indicators;

[0105] The individualized result output module is used to stratify the prediction results into high-risk, medium-risk, and low-risk levels. The output results include: individual recurrence probability value, stratified risk level, and explanations of key driving medical indicators. The prediction results are also pushed to the doctor-side clinical decision support system to assist in formulating personalized postoperative management strategies and follow-up plans.

[0106] A method for predicting the recurrence of common bile duct stones after ERCP based on machine learning is applied to a system for predicting the recurrence of common bile duct stones after ERCP based on machine learning, comprising the following steps:

[0107] S1. Collect multidimensional clinical data sets from patients before, during, and after ERCP to form a clinical data set for prediction;

[0108] S2. Preprocess the clinical dataset and construct a standardized preprocessed clinical dataset;

[0109] S3. Construct feature vectors based on the preprocessed clinical dataset to establish a feature dataset suitable for the input requirements of a bidirectional gated recurrent unit network.

[0110] S4. Build an improved bidirectional gated recurrent unit network model to predict recurrence risk;

[0111] S5. Based on key indicators reflecting the recurrence of common bile duct stones after ERCP in the preprocessed clinical dataset, a medically guided initialization of the hyperparameter search space was performed to form an optimized initial population. A global search for the hyperparameters of the improved bidirectional gated recurrent unit network model was performed using the Grey Wolf Optimization Algorithm. Dynamic fitness evaluation was performed, and the population position was updated to obtain the optimal hyperparameters. This optimal hyperparameter configuration was applied to the improved bidirectional gated recurrent unit network model to obtain the optimized improved bidirectional gated recurrent unit network model.

[0112] S6. Use the optimized improved bidirectional gated recurrent unit network model to predict the recurrence risk of new preprocessed clinical datasets and output individualized prediction results for the recurrence risk of common bile duct stones after ERCP.

[0113] In this embodiment, S1 includes the following steps:

[0114] S11. Set the clinical data collection window to collect the patient's multi-dimensional clinical raw data before, during and after ERCP to form the initial clinical data set :

[0115] ;

[0116] in, Indicates the Clinical data records of patients, is the total number of patients, Indicates the The surgery time of the patients Represents the preoperative data vector, which includes patient demographic information and basic examination indicators. Represents the intraoperative data vector, which includes variables related to surgical records. Represents the postoperative data vector, which includes biochemical indicators, imaging data and follow-up record characteristics;

[0117] S12. For initial clinical data collection A unified feature space mapping function is defined for different types of clinical data. The feature space mapping function includes unified data encoding, time alignment, and structured rearrangement operations. The feature space mapping function maps each type of data to a standardized feature space to form preoperative feature sets, intraoperative feature sets, and postoperative feature sets.

[0118] S13. Integrate the preoperative feature set, intraoperative feature set, and postoperative feature set in chronological order to construct a clinical dataset for modeling. Normalize the timestamp set in the clinical dataset to form a standardized time axis for calibrating the time series alignment operation between patients. and timeline as standardized input.

[0119] In this embodiment, S2 includes the following steps:

[0120] S21. Clinical Datasets The eigenvectors in Perform missing value detection and processing, and use the mean filling strategy to fill in the missing value feature dimensions. The patient's When a feature value is missing, the value of the first The arithmetic mean of the eigenvalues is replaced to form the eigenvector after the missing values are filled;

[0121] S22. Based on the missing value filled feature vector, perform outlier detection and processing. For each feature dimension, calculate the median of all patients' feature values and the absolute value of the deviation from the median. Use this to identify outliers. When the deviation between a feature value and the median exceeds a preset tolerance multiple, replace the feature value with the median of the current feature dimension, forming a feature vector after outlier processing.

[0122] S23. Perform normalization on the feature vector after outlier processing, and standardize all feature values to the interval Within the range, the minimum value of the feature dimension in all patients is used as the lower limit and the maximum value is used as the upper limit. Linear scaling is performed on each feature value to form a normalized feature vector.

[0123] S24. Express all normalized patient feature vectors as the standardized preprocessed clinical feature matrix:

[0124] ;

[0125] in, Indicates the Patient No. The normalized value of the feature, The dimension is The standardized pretreatment clinical characteristics matrix, is the number of patient samples, is the feature dimension;

[0126] S25. Combine the standardized preprocessed clinical feature matrix with the standardized time axis to construct a time-aligned standardized preprocessed clinical dataset .

[0127] In this embodiment, S3 includes the following steps:

[0128] S31. Based on standardized preprocessed clinical dataset , extract the clinical time series feature vector of each patient , combine the feature vectors of all patients in chronological order to construct a two-dimensional input feature dataset, which has OK, The structure of the columns, represents the number of patients, represents the standardized characteristic dimensions of each patient;

[0129] S32. Convert the two-dimensional input feature data set into a sequence format divided by time steps, and set the time step length to , divide the feature vector of each patient into continuous time step input, each time step input contains a normalized feature dimension consistent eigenvectors;

[0130] S33. Establishing a Hyperparameter Search Space for an Improved Bidirectional Gated Recurrent Unit Network Model , the hyperparameter search space includes the number of network layers , the number of hidden units in each layer , learning rate , dropout rate and batch processing scale ,in:

[0131] The number of network layers is a positive integer, which is used to define the depth of the GRU layer in the model;

[0132] The number of hidden units in each layer is a positive integer, which is used to define the number of neurons in each layer of GRU units;

[0133] The learning rate is a continuous real number between 0 and 1, which is used to control the step size of model parameter updates;

[0134] The dropout rate is a continuous real number between 0 and 1, which is used to control the random dropout ratio of nodes in the model;

[0135] The batch size is a positive integer that sets the number of samples used for each round of parameter update in model training.

[0136] S34. For each hyperparameter Set the corresponding initial search range, the range of the number of network layers is set from the minimum number of layers to the maximum number of layers, the range of the number of hidden units in each layer is set from the minimum number of hidden units to the maximum number of hidden units, the range of the learning rate is set from the minimum learning rate to the maximum learning rate, the range of the dropout rate is set from the minimum dropout rate to the maximum dropout rate, and the range of the batch size is set from the minimum batch size to the maximum batch size.

[0137] In this embodiment, S4 includes the following steps:

[0138] S41. Based on preoperative data vector , intraoperative data vector and postoperative data vector , set clinical stage labels for each time segment ,in Indicates the corresponding stage;

[0139] S42. Feature vector for each stage , calculate the changing trend of clinical characteristics between this stage and the previous stage, and form a time series enhanced feature vector ;

[0140] S43. Enhance the time series feature vector with clinical stage labels Input into clinical stage perception gating mechanism and define gating weights for:

[0141] ;

[0142] in, is the Sigmoid function, 、 、 are the weights, stage embedding matrices, and biases learned during model training, The gating weight for each feature dimension represents the importance of this indicator to recurrence risk modeling at this stage;

[0143] S44. Construct a bidirectional gated recurrent unit network with the input being the feature vector after stage regulation :

[0144] ;

[0145] in, It is the Hadamard element-wise product, indicating that only the feature dimensions that are highly correlated with the risk of recurrence are retained in a specific stage after ERCP.

[0146] S45. The characteristic vector after stage regulation As input, the In the network structure composed of layers of bidirectional gated recurrent units, each layer is set hidden units, outputting the forward state and backward state , combined with the temporal significance of post-ERCP recurrence in the postoperative period, a risk-guided bidirectional fusion mechanism is used to define the fusion representation :

[0147] ;

[0148] in, is the risk attention weight;

[0149] S46. Fusion representation of all time step outputs Perform weighted pooling to obtain the patient-level global temporal feature expression vector ;

[0150] S47. Express the global temporal feature vector Input to the fully connected layer and predict the output recurrence risk :

[0151] ;

[0152] in, 、 are weight and bias terms, Indicates the The recurrence risk of each patient;

[0153] S48. During the training process, hyperparameters are used for optimization control:

[0154] Learning rate : Control parameter update speed;

[0155] Dropout rate : Used to randomly drop some neurons in the bidirectional gated recurrent unit layer and the fully connected layer;

[0156] Batch size : Control the number of samples selected in each round of parameter update;

[0157] Number of network layers , the number of hidden units : Controlling the structural depth and feature capacity of bidirectional gated recurrent unit networks.

[0158] In this embodiment, S5 includes the following steps:

[0159] S51. Extracting a set of key medical features significantly associated with the risk of postoperative recurrence of common bile duct stones based on a standardized preprocessed clinical dataset The key medical feature set includes the total number of stones removed during surgery, the change rate of TBiL after surgery, the peak value of ALT after surgery, and the results of the 7-day follow-up after surgery. The feature strength vector is constructed ;

[0160] S52. Based on feature intensity vector , introduce the medical feature-guided initialization mechanism to generate the optimized population initialization offset matrix , and inject the optimized population initialization offset matrix into the generation process of the initial population of the gray wolf algorithm:

[0161] ;

[0162] in, is the bias intensity weight, Initialize the hyperparameters for the standard uniform individual and optimize the population initialization offset matrix Depend on Mapping generation allows the hyperparameters corresponding to high-risk medical features to enter the search space first, Optimizing the population size in the algorithm for gray wolves;

[0163] Optimize population initialization offset matrix The system maps in the following ways:

[0164] First, each medical feature is normalized in its entire sample distribution so that its value falls between 0 and 1 to unify the dimension and avoid the impact of feature scale differences on the initialization offset weight; then, a preset hyperparameter-feature-correlation relationship matrix is defined to clarify the sensitive mapping relationship between different medical features and each hyperparameter in the gray wolf optimization search space. For example:

[0165] When the TBiL change rate is large after surgery, the model learning rate and network depth need to be increased to enhance the learning ability of pathological mutations; when the ALT peak is abnormal, it is recommended to increase the number of hidden units and dropout rate to improve model robustness and control overfitting; when the number of residual stones during surgery is large, the batch processing scale should be appropriately increased so that the model can more stably perceive the "complex input" pattern.

[0166] Finally, the system performs matrix multiplication on the normalized medical feature values and the above mapping relationship to calculate the offset value of each hyperparameter in the initial population, and uses this to construct the optimized population initialization offset matrix , which is used to guide the initial population to concentrate in the direction of potential better solutions in the search space.

[0167] S53. Partition the search space by clinical stage, construct a stage-by-stage convergence strategy, and define the search state label for each individual gray wolf in the population. , dynamically control the update step factor according to the search status label , given the high-risk postoperative period, prioritize maintaining global search capabilities:

[0168] ;

[0169] in, is the maximum number of iterations, is the initial step size control factor, t represents the current iteration index in the gray wolf optimization algorithm;

[0170] S54. Based on the training set and the validation set, for each set of candidate hyperparameters Constructing a risk-sensitive dynamic fitness function:

[0171]

[0172] in, Represents the area under the ROC curve of the model on the validation set, represents the recurrence sensitivity of the model on samples in the postoperative period, is the balance coefficient;

[0173] S55. According to the gray wolf optimization strategy, select the best individual, the second best individual, and the second best individual from the current population, and set their positions respectively. 、 、 , then update the The position of each individual is:

[0174]

[0175] in, is the combined Euclidean distance between the current individual and the three best individuals;

[0176] S56. Repeat steps S53-S55, iteratively perform global search until the maximum number of iterations is reached Or the fitness change is less than the preset threshold The convergence conditions of , and finally the optimal hyperparameter configuration is obtained:

[0177]

[0178] S57. Configure the optimal hyperparameters Applied to the bidirectional gated recurrent unit network structure, the final training task is globally optimized, the optimal hyperparameter configuration is applied to the improved bidirectional gated recurrent unit network model, and the preprocessed clinical dataset is iteratively trained until the training of the improved bidirectional gated recurrent unit network model reaches the preset stopping criterion, and the optimized improved bidirectional gated recurrent unit network model is obtained.

[0179] In this embodiment, S6 includes the following steps:

[0180] S71. Input the preprocessed new clinical data set into the optimized improved bidirectional gated recurrent unit network model, perform forward prediction, and obtain the first Output value of stone recurrence probability for new patients ;

[0181] S72. Output value of stone recurrence probability for new patients , combined with the high-risk index threshold after surgery, to construct an individualized recurrence risk classification standard based on medical rules, and divide each patient into the corresponding risk level;

[0182] S73. Output the probability of stone recurrence for each new patient The key indicator data during and after the operation are correlated and statistically analyzed, and the data are stratified by risk level to form an individualized prediction result of the risk of recurrence of common bile duct stones after ERCP.

[0183] In this implementation, the risk levels include:

[0184] High-risk patients are classified as high risk if they meet any of the following conditions: Predicted probability , and the TBiL change rate within 7 days after surgery , or postoperative ALT peak , or the number of stones remaining during surgery ;

[0185] Patients at intermediate risk are classified as those who meet any of the following conditions: Predicted probability , and the TBiL change rate within 7 days after surgery , or the number of stones remaining during surgery , or postoperative ALT peak ;

[0186] Low-risk patients are classified as low-risk if they meet any of the following conditions: Predicted probability , and the TBiL change rate after surgery , and the peak ALT level after surgery , and there are no residual stones during the operation, that is .

[0187] The present invention constructs data input channels for the preoperative, intraoperative, and postoperative stages, and generates time-series enhancement features based on the changing trend of the feature vectors in each stage. On this basis, a clinical stage-aware gating mechanism is proposed. By introducing label embedding and feature gating weight learning for each stage, the expression intensity of each feature dimension in different time segments is dynamically adjusted, effectively emphasizing the contribution of postoperative biochemical indicators and imaging information to recurrence modeling. The feature flow direction can also be automatically adjusted according to the stage-by-stage risk weight, effectively improving the model's ability to pay attention to recurrence-sensitive areas.

[0188] The present invention embeds key medical features into the initial population construction of the gray wolf optimization algorithm to form a medical feature-driven initialization offset matrix, and introduces a stage-by-stage convergence control strategy to maintain global search capabilities in the early high-risk stage after surgery and quickly converge in the preoperative stage to stabilize the basic network.

[0189] The present invention constructs a risk-guided bidirectional fusion mechanism based on GRU, introduces key stage weights in the network output stage, forms a temporal aggregation representation of dynamic risk perception, and combines weighted pooling to enhance the model's ability to integrate long-term dependent information. In actual deployment, it has significant clinical value for the early identification of high-risk individuals.

[0190] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A machine learning-based system for predicting the recurrence of common bile duct stones after ERCP, characterized in that: Includes the following modules: The data acquisition module is used to collect multidimensional clinical data of patients before, during and after ERCP to form the original clinical data set; A data preprocessing module is used to perform preprocessing operations on the original clinical data set and construct a standardized preprocessed clinical data set; Feature construction and hyperparameter search space definition module, which is used to construct preoperative, intraoperative, and postoperative stage feature vectors based on standardized preprocessed clinical datasets and define the hyperparameter search space of the improved bidirectional gated recurrent unit network model; Improved bidirectional gated recurrent unit network model construction module, used to construct an improved bidirectional gated recurrent unit network model and output a recurrence risk prediction value; The Gray Wolf Optimization hyperparameter tuning module constructs a medically guided initialization population and performs multiple rounds of Gray Wolf optimization searches to obtain the optimal hyperparameter configuration, ultimately generating an optimized improved bidirectional gated recurrent unit network model. The model training and prediction module is used to use the optimized improved bidirectional gated recurrent unit network model to perform forward predictions on standardized new patient clinical data, output recurrence probability values, and construct risk classification rules based on the predicted probability and key postoperative indicators; Individualized result output module, which is used to output prediction results in layers according to high-risk, medium-risk, and low-risk levels, and push the prediction results to the doctor's end; The feature construction and hyperparameter search space definition module includes: Based on the standardized preprocessed clinical dataset D preprocessed , extract the clinical time series feature vector of each patient The feature vectors of all patients are combined in chronological order to construct a two-dimensional input feature dataset. The two-dimensional input feature dataset has a structure of N rows and d columns, where N represents the number of patients and d represents the standardized feature dimension of each patient. Convert the two-dimensional input feature dataset into a sequence format divided by time steps, set the time step length to T, and divide the feature vector of each patient into T consecutive time step inputs according to the timestamp. Each time step input contains a feature vector consistent with the standardized feature dimension d; Establishing a hyperparameter search space Θ for the improved bidirectional gated recurrent unit network model, wherein the hyperparameter search space includes the number of network layers θ1, the number of hidden units per layer θ2, the learning rate θ3, the dropout rate θ4, and the batch size θ5; For each hyperparameter θ j ∈Θ sets the corresponding initial search range; The Gray Wolf optimization hyperparameter tuning module includes: The key medical feature set F that is significantly associated with the risk of recurrence after common bile duct stone surgery was extracted based on the standardized preprocessed clinical data set. risk The key medical feature set includes the total number of stones removed during surgery, the change rate of TBiL after surgery, the peak value of ALT after surgery, and the results of the 7-day follow-up after surgery. The feature strength vector v is constructed med ; Based on the feature intensity vector v med , introduce the medical feature-guided initialization mechanism to generate the optimized population initialization offset matrix Δ i nit, and inject the optimized population initialization offset matrix into the generation process of the initial population of the gray wolf algorithm: Where ω is the offset strength weight, Initialize the hyperparameter individuals uniformly for the standard, and optimize the population initialization offset matrix Δ i nit by v med The mapping is generated so that the hyperparameters corresponding to high-risk medical features are preferentially entered into the search space, and M is the population size in the gray wolf optimization algorithm; Partition the search space by clinical stage, construct a stage-by-stage convergence strategy, and define the search state label for each individual gray wolf in the population. Dynamically control the update step factor based on the search status label Prioritize maintaining global search capabilities during the high-risk postoperative period: Among them, T max is the maximum number of iterations, a0 is the initial step size control factor, and t represents the current iteration index in the gray wolf optimization algorithm; Based on the training set and the validation set, for each set of candidate hyperparameters Constructing a risk-sensitive dynamic fitness function: Among them, AUC val Represents the area under the ROC curve of the model on the validation set, SEN post represents the recurrence sensitivity of the model on samples in the postoperative period, and λ is the balance coefficient; According to the gray wolf optimization strategy, the best individual, the second best individual, and the second best individual are selected from the current population, and their positions are set as follows: Then the position of the mth individual is updated as: in, is the combined Euclidean distance between the current individual and the three best individuals; Repeat the global search until the maximum number of iterations T is reached max Or satisfy the convergence condition that the fitness change is less than the preset threshold ∈, and finally obtain the optimal hyperparameter configuration Θ * ; Configure the optimal hyperparameters Applied to the bidirectional gated recurrent unit network structure, the final training task is globally optimized, the optimal hyperparameter configuration is applied to the improved bidirectional gated recurrent unit network model, and the preprocessed clinical dataset is iteratively trained until the training of the improved bidirectional gated recurrent unit network model reaches the preset stopping criterion, and the optimized improved bidirectional gated recurrent unit network model is obtained.

2. A method for predicting the recurrence of common bile duct stones after ERCP based on machine learning, applied to the system for predicting the recurrence of common bile duct stones after ERCP based on machine learning according to claim 1, characterized in that: The following steps are involved: S1. Collect multidimensional clinical data sets of patients before, during, and after ERCP to form a clinical data set for prediction; S2. Preprocess the clinical dataset and construct a standardized preprocessed clinical dataset; S3. Construct feature vectors based on the preprocessed clinical dataset to establish a feature dataset suitable for the input requirements of the bidirectional gated recurrent unit network; S4. Build an improved bidirectional gated recurrent unit network model to predict recurrence risk; S5. Based on the key indicators reflecting the recurrence of common bile duct stones after ERCP in the preprocessed clinical dataset, perform medical feature-guided initialization of the hyperparameter search space to form an optimized initial population. Use the Gray Wolf Optimization Algorithm to perform a global search for the hyperparameters of the improved bidirectional gated recurrent unit network model. Perform dynamic fitness evaluation, update the population position to obtain the optimal hyperparameters, and apply the optimal hyperparameter configuration to the improved bidirectional gated recurrent unit network model to obtain the optimized improved bidirectional gated recurrent unit network model. S6. Use the optimized improved bidirectional gated recurrent unit network model to predict recurrence risk in a new pre-processed clinical dataset and output individualized post-ERCP common bile duct stone recurrence risk prediction results; The S3 includes the following steps: S31. Based on standardized preprocessed clinical dataset D preprocessed , extract the clinical time series feature vector of each patient The feature vectors of all patients are combined in chronological order to construct a two-dimensional input feature dataset. The two-dimensional input feature dataset has a structure of N rows and d columns, where N represents the number of patients and d represents the standardized feature dimension of each patient. S32. Convert the two-dimensional input feature dataset into a sequence format divided by time steps, set the time step length to T, and divide the feature vector of each patient into T consecutive time step inputs according to the timestamp, where each time step input contains a feature vector consistent with the standardized feature dimension d; S33. Establishing an improved bidirectional gated recurrent unit network model hyperparameter search space Θ, the hyperparameter search space includes the number of network layers θ1, the number of hidden units per layer θ2, the learning rate θ3, the dropout rate θ4 and the batch size θ5; S34. For each hyperparameter θ j ∈Θ sets the corresponding initial search range; The S5 comprises the following steps: S51. Extract key medical feature set F that is significantly associated with the risk of recurrence after common bile duct stone surgery based on standardized preprocessed clinical dataset risk The key medical feature set includes the total number of stones removed during surgery, the change rate of TBiL after surgery, the peak value of ALT after surgery, and the results of the 7-day follow-up after surgery. The feature strength vector v is constructed med ; S52. Based on the feature intensity vector v med , introduce the medical feature-guided initialization mechanism to generate the optimized population initialization offset matrix Δ i nit, and inject the optimized population initialization offset matrix into the generation process of the initial population of the gray wolf algorithm: Where ω is the offset strength weight, Initialize the hyperparameter individuals uniformly for the standard, and optimize the population initialization offset matrix Δ i nit by v med The mapping is generated so that the hyperparameters corresponding to high-risk medical features are preferentially entered into the search space, and M is the population size in the gray wolf optimization algorithm; S53. Partition the search space by clinical stage, construct a stage-by-stage convergence strategy, and define the search state label for each individual gray wolf in the population. Dynamically control the update step factor based on the search status label Prioritize maintaining global search capabilities during the high-risk postoperative period: Among them, T max is the maximum number of iterations, a0 is the initial step size control factor, and t represents the current iteration index in the gray wolf optimization algorithm; S54. Based on the training set and the validation set, for each set of candidate hyperparameters Constructing a risk-sensitive dynamic fitness function: Among them, AUC val Represents the area under the ROC curve of the model on the validation set, SEN post represents the recurrence sensitivity of the model on samples in the postoperative period, and λ is the balance coefficient; S55. According to the gray wolf optimization strategy, select the best individual, the second best individual, and the second best individual from the current population, and set their positions respectively. Then the position of the mth individual is updated as: in, is the combined Euclidean distance between the current individual and the three best individuals; S56. Repeat steps S53-S55, iteratively perform global search until the maximum number of iterations T is reached max Or satisfy the convergence condition that the fitness change is less than the preset threshold ∈, and finally obtain the optimal hyperparameter configuration Θ * ; S57. Configure the optimal hyperparameters Applied to the bidirectional gated recurrent unit network structure, the final training task is globally optimized, the optimal hyperparameter configuration is applied to the improved bidirectional gated recurrent unit network model, and the preprocessed clinical dataset is iteratively trained until the training of the improved bidirectional gated recurrent unit network model reaches the preset stopping criterion, and the optimized improved bidirectional gated recurrent unit network model is obtained.

3. The method for predicting recurrence of common bile duct stones after ERCP based on machine learning according to claim 2, characterized in that: Said S1 comprises the following steps: S11. Set the clinical data collection window to collect the patient's multi-dimensional clinical raw data before, during and after ERCP to form the initial clinical data set D raw : Among them, d i represents the clinical data record of the i-th patient, N is the total number of patients, t i represents the operation time point of the i-th patient, Represents the preoperative data vector, which includes patient demographic information and basic examination indicators. Represents the intraoperative data vector, which includes variables related to surgical records. Represents the postoperative data vector, which includes biochemical indicators, imaging data and follow-up record characteristics; S12. For the initial clinical data set D raw A unified feature space mapping function is defined for different types of clinical data. The feature space mapping function includes unified data encoding, time alignment, and structured rearrangement operations. The feature space mapping function maps each type of data to a standardized feature space to form preoperative feature sets, intraoperative feature sets, and postoperative feature sets. S13. Integrate the preoperative feature set, intraoperative feature set, and postoperative feature set in chronological order to construct a clinical dataset for modeling. Normalize the timestamp set in the clinical dataset to form a standardized time axis for calibrating the time series alignment operation between patients. clinical and time axis t i as standardized input.

4. The method for predicting recurrence of common bile duct stones after ERCP based on machine learning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Clinical Datasets The eigenvector x in i Missing value detection and processing are performed. For feature dimensions with missing values, the mean filling strategy is used to fill them. When the j-th feature value of the i-th patient is missing, the arithmetic mean of the j-th feature value of all patients who do not have this feature missing is used to replace it to form the feature vector after missing value filling. S22. Based on the missing value filled feature vector, perform outlier detection and processing. For each feature dimension, calculate the median of all patients' feature values and the absolute value of the deviation from the median. Use this to identify outliers. When the deviation between a feature value and the median exceeds a preset tolerance multiple, replace the feature value with the median of the current feature dimension, forming a feature vector after outlier processing. S23. Perform a normalization operation on the feature vector after outlier processing, normalizing all feature values to the range of [0, 1], using the minimum value of the feature dimension across all patients as the lower limit and the maximum value as the upper limit, and performing linear scaling on each feature value to form a normalized feature vector; S24. Represent all normalized patient feature vectors as the standardized preprocessed clinical feature matrix X norm ; S25. Combine the standardized preprocessed clinical feature matrix with the standardized time axis to construct a time-aligned standardized preprocessed clinical dataset 5. The method for predicting recurrence of common bile duct stones after ERCP based on machine learning according to claim 2, characterized in that: The S4 comprises the following steps: S41. Based on preoperative data vector Intraoperative data vector and postoperative data vector Set clinical stage labels for each time segment Where s∈{pre,intra,post} represents the corresponding stage; S42. Feature vector for each stage Calculate the changing trend of clinical characteristics between this stage and the previous stage to form a time series enhanced feature vector S43. Enhance the time series feature vector with clinical stage labels Input into clinical stage perception gating mechanism and define gating weights for: Where σ(·) is the Sigmoid function, W g 、U g 、b g are the weights, stage embedding matrices, and biases learned during model training, The gating weight for each feature dimension represents the importance of this indicator to recurrence risk modeling at this stage; S44. Construct a bidirectional gated recurrent unit network with the input being the feature vector after stage regulation Among them, ⊙ is the Hadamard element-by-element product, which means that only the feature dimensions that are highly correlated with the risk of recurrence are retained in a specific stage after ERCP. S45. The characteristic vector after stage regulation As input, it is sent to the network structure composed of θ1 layers of bidirectional gated recurrent units, each layer has θ2 hidden units, and outputs the forward state and backward state Combined with the temporal significance of post-ERCP recurrence in the postoperative period, a risk-guided bidirectional fusion mechanism was used to define the fusion representation h t ; S46. The fusion representation h of all time step outputs t Perform weighted pooling to obtain the patient-level global temporal feature expression vector H final ; S47. Express the global temporal feature vector H final Input to the fully connected layer and predict the output recurrence risk 6. The method for predicting recurrence of common bile duct stones after ERCP based on machine learning according to claim 2, characterized in that: The S6 comprises the following steps: S61. Input the preprocessed new clinical data set into the optimized improved bidirectional gated recurrent unit network model, perform forward prediction, and obtain the output value of the stone recurrence probability of the jth new patient. S62. Output value of stone recurrence probability for new patients Combined with the high-risk index threshold after surgery, an individualized recurrence risk classification standard based on medical rules is constructed, and each patient is divided into the corresponding risk level; S63. Output the probability of stone recurrence for each new patient The key indicator data during and after the operation are correlated and statistically analyzed, and the output is stratified by risk level to form an individualized prediction result for the recurrence risk of common bile duct stones after ERCP.

7. The method for predicting recurrence of common bile duct stones after ERCP based on machine learning according to claim 2, characterized in that: The risk levels include: High-risk patients are classified as high risk if they meet any of the following conditions: Predicted probability And the TBiL change rate within 7 days after surgery or postoperative ALT peak or the number of residual stones during surgery Patients at intermediate risk are classified as those who meet any of the following conditions: Predicted probability And the TBiL change rate within 7 days after surgery or the number of residual stones during surgery or postoperative ALT peak Low-risk patients are classified as low-risk if they meet any of the following conditions: Predicted probability And the TBiL change rate after surgery And the peak value of ALT after surgery And there are no residual stones during the operation, that is

Images