基于血清微小核糖核酸与临床因素的食管癌内镜黏膜下剥离术后复发预测模型OA

Objective To explore the predictive value of machine learning models for predicting the recurrence risk after endoscopic submucosal dissection(ESD)in patients with early-stage esopha-geal cancer,which were constructed by integrating clinical factors and serum microRNA-204(miR-204)and microRNA-134(miR-134).Methods Early-stage esophageal cancer patients were selected as the study subjects in a 1∶1 ratio,including 100 patients with recurrence within 1 year af-ter ESD and 100 patients without recurrence,who were respectively included in recurrence group and non-recurrence group.The clinical data of the two groups were compared.The least absolute shrink-age and selection operator(LASSO)regression model was used to screen variables,and four machine learning models,namely random forest(RF),logistic regression(LR),extreme gradient boosting(XGBoost),and support vector machine(SVM),were constructed.The receiver operating characteris-tic(ROC)curve was plotted to analyze the predictive efficacy of the four machine learning models for recurrence after ESD in early-stage esophageal cancer.Additionally,50 patients with recurrence and 50 patients without recurrence after ESD for early-stage esophageal cancer were selected as external validation set to validate the model with the best comprehensive predictive performance through cali-bration curve and decision curve analysis(DCA).Results There were statistically significant differences between the two groups in terms of lesion length,depth of lesion invasion,circumferen-tial extent of the lesion,positive resection margin,and serum levels of miR-204 and miR-134(P＜0.05).Among the four machine learning models,the RF model had the highest F1 score and area under the curve(AUC)for predicting postoperative recurrence,demonstrating the best comprehen-sive predictive performance.In the RF model,the important feature variables were ranked in the fol-lowing order:circumferential extent of the lesion,miR-204,miR-134,positive resection margin,depth of lesion invasion,and lesion length.In the external validation set,the C-index of the RF mod-el was 0.892,and the Brier score was 0.112.The calibration curve and DCA curve showed that the predicted values of the RF model were close to the actual incidence rate,and it had a significant net benefit when the risk threshold range was 15％to 100％.Conclusion Among the four machine learning models constructed based on circumferential extent of the lesion,miR-204,miR-134,positive resection margin,depth of lesion invasion,and lesion length,the RF model has the best comprehen-sive predictive performance for the recurrence risk after ESD in early-stage esophageal cancer,with high discrimination,accuracy,and good clinical applicability.