A clinical-EUS model for GIST probability in resection-selected gastric subepithelial lesions: development and external validation

Abstract Preoperative distinction of gastrointestinal stromal tumors (GISTs) from other gastric subepithelial lesions remains difficult. We evaluated a three-variable clinical and endoscopic ultrasound (EUS) model for GIST probability in patients undergoing endoscopic resection. This retrospective study included 546 development patients. The three-variable logistic model used age, fundic location, and muscularis propria origin in 528 patients with evaluable predictors. Internal validation used 2000 bootstrap resamples. The original coefficients were tested without refitting in 80 patients from a second hospital in the same city (34 GISTs). Apparent and optimism-corrected areas under the receiver operating characteristic curve were 0.805 and 0.803. External discrimination was 0.783 (95% CI 0.680–0.885), with Brier score 0.192, calibration intercept − 0.339 (95% CI − 0.849 to 0.172), and slope 0.823 (95% CI 0.408–1.237). Adding tumor size > 15 mm increased AUC by 0.007 (P = 0.133). In this small external cohort, the model retained moderate discrimination but calibration was uncertain. It may help estimate GIST probability when EUS wall-layer origin is clearly assigned, but it should not be used alone to determine surveillance, biopsy, or resection.