Background and aimsNon-tunneling and submucosal tunneling endoscopic resection (STER) techniques are the most frequent treatments for cardial submucosal tumor (SMT). Here, we analyzed common machine learning (ML) algorithms and compared them with traditional regression models in surgical decision-making for cardial SMTs. MethodsUsing key baseline predictive factors, ML algorithms and logistic regression (LR) were conducted in 246 patients. For the ML algorithms, gradient boosting machines (GBM), artificial neural networks (ANN), random forests (RF), and support vector machines (SVM), were included. For small sample-sized data, a technique for k-fold cross-validation was exploited to avoid over-fitting. Meanwhile, we tuned the parameters through several replications. Then, we quantified the discrimination (area under the curve, AUC) and predictive ability (Brier score, F1 score, specificity, sensitivity, and accuracy) of models. We divided patients (n = 246) into STER-treated (n = 97) and non-tunneling endoscopic resection (NTER)-treated (n = 149) groups. ResultsLR outperformed among all groups (Brier score = 0.1398, F1 score = 0.7391, AUC = 0.8729, and predictive accuracy = 80.65 %). In comparison to ML algorithms, an outperformance of the traditional regression approach was also found in a low-dimensional setting for surgical decision prediction of cardial SMTs. ConclusionsThe traditional regression approach outperformed ML algorithms for the prediction of the best surgical method in patients with SMTs.
Read full abstract