Prediction of overall and cancer-specific survival in postoperative patients with gastrointestinal stromal tumors based on explainable models

Postoperative gastrointestinal stromal tumor (GIST) patients exhibit heterogeneous survival outcomes. Traditional prognostic tools fail to achieve individualized time-specific survival prediction. This study aimed to construct explainable models to predict overall survival (OS) and cancer-specific survival (CSS) for postoperative GIST patients. A total of 3,511 GIST patients from the SEER database were randomly allocated into training and internal validation cohorts (7:3). Additionally, 219 patients from a single medical center served as the external validation cohort. Key variables were screened via three feature selection methods. Four survival models, including Random Survival Forest (RSF), GBM, CoxBoost and DeepSurv, were established. Multiple quantitative metrics and SHAP algorithm were utilized for model evaluation and interpretation. Different predictive variables were included in the final OS and CSS models. The RSF model presented the most stable predictive performance. In internal validation, its C-index for OS and CSS reached 0.735 and 0.776, with favorable time-dependent AUC values. External validation further verified its reliable discrimination and calibration. Age predominantly affected OS, while tumor size and mitotic rate dominated CSS prediction. An online individualized survival calculator was developed based on RSF. The RSF model showed relatively stable predictive performance and provided interpretable predictions. The established online calculator can serve as a convenient clinical tool for postoperative risk stratification and individualized prognostic evaluation of GIST patients.