Development and Evaluation of Machine Learning Models for Predicting Liver Metastasis in Gastrointestinal Stromal Tumours

Gastrointestinal stromal tumours (GISTs) are the most common mesenchymal neoplasms of the gastrointestinal tract and frequently metastasise to the liver, which is associated with unfavourable clinical outcomes. Although contrast-enhanced computed tomography remains the standard modality for detecting liver metastasis, identifying patients with an increased likelihood of metastatic disease may facilitate individualised risk stratification and optimise clinical management in situations where complete staging information is temporarily unavailable. In this retrospective study, data from 5,206 patients diagnosed with GIST were extracted from the Surveillance, Epidemiology, and End Results (SEER) database. The primary endpoint was the presence of liver metastasis at diagnosis. Multiple machine-learning models, including Random Forest (RF), Support Vector Machine (SVM), Feature-Weighted Support Vector Machine (FWSVM), and Artificial Neural Network (ANN), were developed and evaluated for liver metastasis prediction. Model performance was assessed using accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUROC). Model interpretability was evaluated using SHAP (SHapley Additive exPlanations), and Cox proportional hazards analysis was performed to assess the prognostic significance of liver metastasis. Among the evaluated models, the ANN achieved the best overall predictive performance, with an AUROC of 0.862 (95% CI: 0.832-0.890) and a PR-AUC of 0.60 in the independent testing cohort. Tumour size, pathological grade, and primary tumour site were consistently identified as the most influential predictors. SHAP analyses provided both global and local explanations of model predictions, improving model interpretability. Survival analyses demonstrated that patients with liver metastasis had significantly worse overall survival than those without liver metastasis. These findings suggest that machine-learning approaches may serve as complementary tools for liver metastasis risk stratification and for identifying clinico pathological factors associated with liver metastasis at diagnosis. External validation and prospective clinical evaluation are warranted before routine clinical implementation.