Development and validation of a machine-learning model for preoperative risk of gastric gastrointestinal stromal tumors

Abstract

Background

Gastrointestinal stromal tumors (GISTs) have malignant potential, and treatment varies according to risk. However, no specific protocols exist for preoperative assessment of the malignant potential of gastric GISTs (gGISTs). This study aimed to use machine learning (ML) to develop and validate clinically relevant preoperative models to predict the malignant potential of gGISTs.

Methods

This study screened patients diagnosed with gGISTs at the Affiliated Hospital of North Sichuan Medical College. Moreover, this study employed the Least Absolute Shrinkage and Selection Operator (LASSO) and logistic regression to identify risk factors. Subsequently, an ensemble of ML models was used to determine the optimal classifier. In addition, this study used SHapley Additive exPlanations (SHAP) for tailored risk profiling.

Results

This study included 318 patients with gGISTs. Using LASSO regression and multifactorial logistic regression, this study analyzed the training dataset, revealing that the presence of endoscopic ultrasound (EUS) high-risk features, tumor border clarity, tumor diameter, and monocyte-to-lymphocyte ratio (MLR) were significant predictors of high malignancy risk in gGIST. As determined by our ML approach, the logistic classification model demonstrated optimal performance, with area under the receiver operating characteristic curves of 0.919 for the training set and 0.925 for the test set. Furthermore, decision curve analysis confirmed the clinical relevance of the model.

Conclusion

High-risk EUS features, ill-defined tumor margins, larger tumor diameters, and elevated MLR independently predicted increased malignant potential in gGIST. This study developed logistic regression models based on these factors, which were further interpreted using the SHAP methodology. This analytical approach facilitated personalized therapeutic decision-making among diverse patient populations.