| Background and Objective: The early clinical signs of malignant obstructive jaundice(MOJ)are atypical and most patients present at an early stage,missing the best time for surgery.The early diagnosis of MOJ patients is still a major challenge.This study aims to use machine learning to construct interpretable predictive models for early and rapid identification of patients with MOJ and to optimise the management of patients with outpatient obstructive jaundice.Methods: Clinical data were retrospectively collected from June 2016 to December 2021 from patients hospitalized at the First Hospital of Lanzhou University and identified as having obstructive jaundice.Patients were randomly divided into training and validation cohorts in a 7:3 ratio;the sample size in the training cohort was balanced using the synthetic minority over-sampling technique(SMOTE),and the predictors of MOJ were analyzed using the Lasso regularization technique.Based on machine learning algorithms to build Light GBM,extreme gradient boosting(XGBoost),support vector machine(SVM),random forest(RF),logistic regression(LR)five kinds of prediction models.Sensitivity,Specificity,Accuracy,positive predictive value(PPV),negative predictive value(NPV),F1 score and other performance parameters of the predictive model were calculated.At the same time,the receiver operating characteristic curve(ROC)and area under ROC curve(AUC)were combined to comprehensively evaluate the model,and the model with the best predictive performance was finally selected.Calibration Plot and Decision Curve Analysis(DCA)were used to assess the calibration and clinical applicability of the prediction models,and SHapley Additive ex Planations(SHAP)was used to visualize the interpretation of the best model.Results: A total of 1062 patients with obstructive jaundice ranging in age from19 to 89 years were enrolled in this study.There were 622 males and 440 females,289 patients with MOJ,accounting for 27.3% of the total sample size,and 773 patients with Benign Obstructive Jaundice(BOJ),accounting for 72.7% of the total sample size.The total sample size was randomly divided into a training cohort(743 cases)and a verification cohort(319 cases)at a ratio of 7:3.In the training cohort,203 cases of MOJ and 540 cases of BOJ were found.SOMTE method was used to balance the sample size in the training cohort.Based on the SMOTE training cohort,age,BMI,red blood cell,platelet,neutrophil ratio,alanine aminotransferase,alkaline phosphatase,total bilirubin,uric acid,triglyceride,low density lipoprotein cholesterol,carbohydrate antigen 199 were determined by Lasso regression method as effective predictors of MOJ.Feature importance ranking showed that total bilirubin was the most important predictor of MOJ.The above features were included in the construction of five machine learning models,and the generalization ability of Light GBM model was tested by validation queue.The results showed that Lightg BM model had the best predictive performance,and its AUC in validation queue was0.893(95%CI: 0.841-0.922).The calibration curve shows that the actual probability of Light GBM model in the verification queue is in good agreement with the predicted probability.The decision curve shows that the model has a good clinical net benefit in most threshold probability ranges.Conclusion: Based on Lasso regularization technique,age,BMI,red blood cell,platelet,neutrophil ratio,alanine aminotransferase,alkaline phosphatase,total bilirubin,uric acid,triglyceride,low density lipoprotein cholesterol and carbohydrate antigen 199 were identified as important predictors of MOJ recognition.The Light GBM model built with the above features has good predictive performance.The visualization of the model can assist medical staff to quickly identify the high-risk population of MOJ in the early stage,and provide scientific suggestions for the next diagnosis and treatment strategy.It can be used as an early differential diagnosis tool for patients with MOJ in the outpatient department. |
PDF Full Text Request