A machine learning method for early prediction of biliary atresia based on routine examination

Abstract

The present invention relates to a machine learning method for early prediction of biliary atresia based on routine examination. At present, the gold standard for BA diagnosis is intraoperative cholangiography. This is an invasive surgery with great technical difficulty and is limited to use in large medical centers. The present invention includes data collection and data preprocessing, and the data is saved as a database file, and also includes model training of the data. The model training is to collect five important feature data, namely γ-glutamyl transferase, liver portal fibrous mass, body weight, gallbladder abnormality and stool color, and then four models including support vector machine (SVM), logistic regression (LR), random forest RF and gradient boosted decision tree (GBDT) are integrated into a set, named biliary atresia prediction model (BAPM), and the predicted risk of biliary atresia is obtained by the weighted voting process of the output probability of the four models. The present invention uses routine examinations to help patients predict biliary atresia non-invasively at an early stage and promote clinical decision-making.

Description

Technical Field

[0001] The present invention relates to a machine learning method for early prediction of biliary atresia based on routine examinations, which screens and utilizes important features to help medical professionals more effectively predict biliary atresia, thereby facilitating clinical decision-making. Background Art

[0002] Biliary atresia (BA) is a serious neonatal disease that can lead to cholestasis and progressive liver failure. BA occurs in approximately 1 in 8,000 to 18,000 live births and, without intervention, will result in death within the first two years of life. Infant prognosis depends largely on early referral and timely implementation of Kasai portoenterostomy. When Kasai portoenterostomy is performed within 45 days of life, survival rates can reach 65.5% at 2 years of age and 40.5% at 15 years of age, respectively. Therefore, early diagnosis is important. Currently, the gold standard for the diagnosis of BA is intraoperative cholangiography. This is an invasive procedure that is technically difficult and is limited to large medical centers.

[0003] It is important to develop strategies for early detection of BA. In some countries and regions, stool color distribution charts have become widely available, which helps to speed up early referral and reduce the time to Kasai portoenterostomy. However, the sensitivity of stool color card screening was only 76.5% at a one-month examination. In a US cohort study, direct / conjugated serum bilirubin measurement was used for BA screening, but the sensitivity estimates had wide 95% confidence intervals (CIs), and these screening tools could not guarantee complete identification of false-negative results. γ-Glutamyltransferase (GGT) is the most widely used biomarker for early diagnosis of BA, but its diagnostic efficacy in different subpopulations is questionable. Researchers have also identified several new potential biomarkers that may be helpful in the diagnosis of biliary atresia, such as circulating microRNAs, matrix metalloproteinase-7, fecal protein, interleukin 33, Th17-related cytokines, urine metabolomics, anti-smooth muscle antibodies, heat shock protein 90, and positive bile duct epithelial cells CD56. However, these tests may be impractical in countries and regions with underdeveloped medical infrastructure. In addition, the echogenic fibrous tissue in front of the portal vein (called the triangular cord sign) detected by B-ultrasound is a very important imaging indicator, but this depends on the level of the operator, with a sensitivity ranging from 49% to 73%. Summary of the invention

[0004] In view of the above problems, the present invention provides a machine learning method for early prediction of biliary atresia based on routine examinations, which uses routine examinations to help patients predict biliary atresia non-invasively at an early stage and achieve timely referral.

[0005] To this end, the present invention adopts the following technical scheme: a machine learning method for early prediction of biliary atresia based on routine examination, including data collection and data preprocessing, saving the data into a database file, characterized in that it also includes model training of the data, the model training is to collect five important feature data, namely γ-glutamyl transferase, portal fibrous mass, body weight, gallbladder abnormality and stool color, and then integrate four models including support vector machine (SVM), logistic regression (LR), random forest RF and gradient boosted decision tree (GBDT) into a set, named biliary atresia prediction model (BAPM), the predicted risk of biliary atresia is obtained through a weighted voting process of the output probabilities of the four models.

[0006] Preferably, the weights of each model in the weighted voting, SVM, LR, RF and GBDT, are 0.2, 0.2, 0.35 and 0.25 respectively, wherein the prediction probability of the biliary atresia prediction model (BAPM) lower than 0.58 is determined to be a low risk of biliary atresia, and the output equal to or higher than 0.58 is predicted to be a high risk of biliary atresia.

[0007] Preferably, the data is divided into a training set and a test set. The test set is data that does not participate in model training and is used to simulate the performance of the model in actual applications.

[0008] Preferably, the five important feature data are selected by LASSO linear regression. LASSO (least absolute shrinkage and selection operator) linear regression is a linear regression method using L1 norm as a regularization term. In LASSO regression, the loss function includes two parts: a data fitting term (i.e., least squares error) and a regularization term of L1 norm. L1 norm refers to the sum of the absolute values ​​of each element in a vector, and is used to measure the number of non-zero elements in a vector. In LASSO regression, by introducing the L1 norm regularization term, the weights of some features will tend to zero during the training process of the model, thereby achieving the sparsity of the features. In other words, LASSO achieves the effect of feature selection by shrinking the weights of some unimportant features to zero.

[0009] Preferably, the data collection is to obtain medical records and clinical data through an electronic health record system (EHR), obtain real-time data through medical equipment, or obtain laboratory result data from a laboratory database. Collect relevant information of patients. The corresponding data can be obtained under the premise of ensuring that legal data acquisition rights are obtained and complying with medical data protection regulations.

[0010] Preferably, the data preprocessing includes deleting useless data, data cleaning, standardization and error correction for the received data.

[0011] Data preprocessing is an important step that aims to clean, transform, and prepare raw data so that it can better suit the requirements of machine learning algorithms.

[0012] The present invention is a multivariate screening model based on conventional non-invasive testing, which is helpful for the early identification of biliary atresia and promotes clinical decision-making. The BAPM model developed by the present invention obtained an AUC of 0.93439 in the internal validation cohort XHV, with a 95% confidence interval (CI) of 0.8927-0.9685. An AUC of 0.94969 was obtained in the external validation cohort CH, with a 95% confidence interval (CI) of 0.9220-0.9725. This shows that the BAPM model developed by us successfully achieves accurate prediction of biliary atresia in infants. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the learning prediction and evaluation process of the present invention.

[0014] Figure 2 This is a schematic diagram of the division of patient cohorts in the present invention.

[0015] Figure 3 Schematic diagram of F1 score determination for the ensemble model BAPM at different cutoff thresholds in the training cohort.

[0016] Figure 4 Schematic diagram of selection of five important characteristic data of the present invention.

[0017] Figure 5 Schematic diagram of ROC curves of BAPM using five significant features in cohorts (a) XHV and (b) CH. DETAILED DESCRIPTION

[0018] The present invention is further described below in conjunction with specific embodiments.

[0019] Figure 1 The learning prediction and evaluation process diagram of the present invention includes data collection and data preprocessing, saving the data into a database file, dividing the data into a training set and a test set, training the model, and evaluating the obtained model after the training is completed. After the training is completed, it is very important to evaluate the obtained model in order to understand the performance and generalization ability of the model. In this embodiment, the original data set is divided into a training set and a test set. The training set is used to train the model, and then the test set is used to evaluate the performance of the model. Specifically, various indicators (such as accuracy, precision, recall rate, F1 score, etc.) can be used to measure the performance of the model on the test set.

[0020] The data collection mentioned above is to obtain medical records and clinical data through the electronic health record system (EHR), obtain real-time data through medical equipment, or obtain laboratory result data from the laboratory database. Collect relevant information of patients. The relevant data can be obtained on the premise of ensuring the legal data acquisition authority and complying with medical data protection regulations.

[0021] In data preprocessing, useless data indicators include name, hospital number, etc., which belong to diagnostic data format indicators. They can indicate the source of sample data, etc., but have no effect on machine learning prediction, so the corresponding data is deleted.

[0022] Data cleaning is to deal with some errors or anomalies in the data, including: Handling missing values: Detect and process missing values ​​in the data. You can delete the rows or columns where the missing values ​​are located, or use interpolation and other methods to fill in the missing values. Features with missing values ​​accounting for 5% or more of the total value are discarded, such as bile acid. On the contrary, for features with missing values ​​less than 5%, data interpolation is used to complete it. For example, for alkaline phosphatase, we use the mean value for interpolation, and for GGT, median interpolation is used to enhance robustness to outliers. Subsequently, in order to mitigate the impact of different scales and prevent weight imbalance in feature selection, we normalize the continuous features to the range of 0 to 1.

[0023] Handling outliers: Detecting and handling outliers in data can be done by using statistical methods or outlier detection algorithms to identify and handle outliers.

[0024] Handling duplicate values: Detect and handle duplicate values ​​in data, and remove duplicates based on specific attributes or entire rows of data.

[0025] Standardization is to process data into a format that is convenient for machine learning, including: feature encoding: converting non-numerical features into numerical representations, feature scaling: scaling numerical features, common methods include normalization and scaling. For example, data features can be scaled to within 0-1 to facilitate unified processing. Numericalization: converting non-numerical data into numerical values ​​for easy processing, such as converting severity levels of low, medium, and high into 1, 2, 3, etc.

[0026] The specific implementation example of model training and verification of data is as follows:

[0027] A total of 872 consecutive cholestatic patients with known outcomes (BA or infantile hepatitis) from a large number of cholestatic infants admitted to Shanghai Xinhua Hospital (XH) and Children's Hospital (CH) were collected, and a total of 79 patients were excluded due to missing data. Finally, 793 cholestatic patients were eligible for inclusion, such as Figure 2As shown in Figure 2, 66.7% of the participants from XH were randomly assigned to the training cohort (XHT cohort), while the remaining 33.3% were assigned to the internal validation cohort (XHV cohort). In addition, infant patients from CH were used as an external validation cohort (CH cohort).

[0028] Among the 370 training samples of the Xinhua Hospital training cohort, after feature selection using the selection operator Lasso, five important features, namely abnormal gallbladder, portal fibrous mass, stool color, body weight and gamma-glutamyl transferase (GGT), were used for disease prediction, where body weight and GGT were normalized to between [0,1].

[0029] Use support vector machine (SVM) to classify new samples. The classification idea of ​​SVM is to construct a hyperplane to separate samples of different categories. The decision function of SVM is defined as

[0030]

[0031] Among them, x is the sample to be classified, N is the number of support vectors, and α i is the weight of the support vector (Lagrange multiplier), y i is the training sample label of the corresponding support vector, and b is the bias term of the support vector model. The decision function uses the weights of the support vectors in the trained SVM model, the labels of the support vectors, the kernel function, and the bias term to calculate the value of the decision function. If this value is greater than zero, the new sample is classified as a positive class; if it is less than zero, it is classified as a negative class.

[0032] For example, for the samples of the Xinhua Hospital training cohort, the bias term b of the trained SVM model is 14.8776, and the total number of support vectors is 153, of which there are 75 samples with a label of -1 and 78 samples with a label of 1.

[0033] The establishment of Logistics regression model (LR). The model establishment process is divided into model training and evaluation. The formula is:

[0034]

[0035] Among them, P is the probability that the output variable takes the value of 1 given the input variable x, that is, the probability of predicting biliary atresia.

[0036] The establishment of the random forest model (RF). The number of decision trees was determined to be 70 through ten-fold cross validation. The process of random forest construction is as follows: randomly select a few features from multiple clinical medical features to determine the decision result of a node on the decision tree; sample multiple times from all training samples with replacement to form a training set, and use the unsampled training samples for prediction; for each node, randomly select several features and calculate the best splitting method; each decision tree is fully grown without pruning to construct a decision tree; repeat the above steps to construct multiple decision trees until the number of decision trees reaches the preset number.

[0037] The establishment of the Gradient Boosted Decision Tree model (GBDT). The process is as follows: In the training phase, GBDT first selects a loss function (mean square error); then fits an initial prediction function, that is, uses the average value of all training samples; in each round of training, GBDT calculates the residual of each sample under the current model, and then uses these residuals as new labels to train the next decision tree; after multiple rounds of iterations, GBDT adds up the prediction results of all decision trees to get the final prediction result.

[0038] In order to exploit the diversity and advantages among different models, we integrated four well-performing models (support vector machine (SVM), logistic regression (LR), random forest (RF), and gradient boosted decision tree (GBDT)) into a machine learning model named Biliary Atresia Prediction Model (BAPM). The predicted risk of biliary atresia determined by our model is obtained through a weighted voting process using the output probabilities of the four models. In this scheme, the assigned individual weights SVM, LR, RF, and GBDT were manually set to 0.2, 0.2, 0.35, and 0.25, respectively. The predicted biliary atresia labels were assigned in the XHT cohort by optimizing the F1 score with a threshold of 0.58, as shown in Figure 3 shown.

[0039] After obtaining the integrated model, we performed biliary atresia risk prediction in the internal validation (XHV) and external testing (CH) cohorts, where the model's predicted probability lower than 0.58 was identified as low risk, while the output equal to or higher than 0.58 was predicted as high risk of biliary atresia.

[0040] LASSO (Least Absolute Shrinkage and Selection Operator) linear regression is a linear regression method that uses the L1 norm as a regularization term. In LASSO regression, the loss function consists of two parts: the data fitting term (i.e., the least squares error) and the L1 norm regularization term. The L1 norm refers to the sum of the absolute values ​​of each element in a vector, which is used to measure the number of non-zero elements in the vector. In LASSO regression, by introducing the L1 norm regularization term, the weights of some features will tend to zero during the training process of the model, thereby achieving feature sparsity. In other words, LASSO achieves the effect of feature selection by shrinking the weights of some unimportant features to zero.

[0041] Figure 4 The specific feature weight value of 0.010 is given to different features with the best feature. The larger the absolute value of the weight, the greater the significance of the corresponding feature in predicting the target variable. As a result, 6 features are colored in gray, and their corresponding weights reach zero, which will be considered as redundant features.

[0042] To reduce computational cost and improve interpretability for clinicians, we increased the regularization coefficient in Lasso to select five features corresponding to the features with the highest absolute value Figure 4 The weights in the test were gamma-glutamyl transferase (GGT), portal fibrous mass (afp), body weight, abnormal gallbladder (abg), and stool color (sc).

[0043] like Figure 5 As shown, the BAPM model developed in the present invention obtained an AUC of 0.93439 in the internal validation cohort XHV, a 95% confidence interval (CI) of 0.8927-0.9685, an accuracy of 90.38%, a sensitivity of 89.52%, a specificity of 87.65%, and an F1 score of 0.8995.

[0044] The AUC of 0.94969 was obtained in the external validation cohort CH, with a 95% confidence interval (CI) of 0.9220-0.9725. In addition, its accuracy, sensitivity, specificity, and F1 score were 86.73%, 86.73%, 87.90%, and 0.8673, respectively. This indicates that the BAPM model we developed successfully achieved accurate prediction of biliary atresia in infants.

[0045] It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope limited by the appended claims of the application.

Claims

1. A machine learning method for early prediction of biliary atresia based on routine examination, including data collection and data preprocessing, saving the data into a database file, characterized in that The method also includes model training on the data, wherein the model training collects five important feature data, namely, γ-glutamyl transferase, hepatic portal fibrous mass, body weight, gallbladder abnormality and stool color, and then integrates four models including support vector machine (SVM), logistic regression (LR), random forest RF and gradient boosted decision tree (GBDT) into a set named as biliary atresia prediction model (BAPM), and the predicted risk of biliary atresia is obtained through a weighted voting process of the output probabilities of the four models; The support vector machine (SVM) is used to classify new samples. The classification idea of ​​SVM is to construct a hyperplane to separate samples of different categories. The decision function of SVM is defined as in, is the sample to be classified, is the number of support vectors, is the weight of the support vector, is the training sample label of the corresponding support vector, is the bias term of the support vector model. The decision function uses the weights of the support vectors in the trained SVM model, the labels of the support vectors, the kernel function, and the bias term to calculate the value of the decision function. If this value is greater than zero, the new sample is classified as a positive class; if it is less than zero, it is classified as a negative class. The process of establishing a logistic regression (LR) model is divided into model training and evaluation, and its formula is: Where P is the probability that the output variable takes the value of 1 given the input variable x, that is, the probability of predicting biliary atresia; The establishment of the random forest RF model is to randomly select a few features from multiple clinical medical features to determine the decision result of a node on the decision tree; sample multiple times with replacement sampling from all training samples to form a training set, and use the unsampled training samples for prediction; for each node, randomly select several features and calculate the best splitting method; each decision tree is fully grown without pruning to construct a decision tree; repeat the above steps to construct multiple decision trees until the number of decision trees reaches a preset value; The establishment process of the gradient boosted decision tree model (GBDT) is as follows: in the training stage, GBDT first selects a loss function; then fits an initial prediction function, that is, uses the average value of all training samples; in each round of training, GBDT calculates the residual of each sample under the current model, and then uses these residuals as new labels to train the next decision tree; after multiple rounds of iterations, GBDT adds up the prediction results of all decision trees to obtain the final prediction result.

2. A machine learning method for early prediction of biliary atresia based on routine examination according to claim 1, characterized in that The weights of support vector machine (SVM), logistic regression (LR), random forest RF and gradient boosted decision tree (GBDT) in the weighted voting were 0.2, 0.2, 0.35 and 0.25, respectively, among which the predicted probability of biliary atresia prediction model (BAPM) lower than 0.58 was determined to be a low risk of biliary atresia, and the output equal to or higher than 0.58 was predicted to be a high risk of biliary atresia.

3. A machine learning method for early prediction of biliary atresia based on routine examination according to claim 1, characterized in that The data is divided into a training set and a test set. The test set is data that does not participate in model training and is used to simulate the performance of the model in actual applications.

4. A machine learning method for early prediction of biliary atresia based on routine examination according to any one of claims 1 to 3, characterized in that The five important feature data are selected through LASSO linear regression.

5. The machine learning method for early prediction of biliary atresia based on routine examination according to claim 1, characterized in that The data collection is to obtain medical records and clinical data through the electronic health record system (EHR), real-time data through medical devices, or laboratory result data obtained from laboratory databases.

6. A machine learning method for early prediction of biliary atresia based on routine examination according to claim 1, characterized in that The data preprocessing includes deleting useless data, cleaning data, standardizing data and correcting errors in the received data.

Images