Preoperative Prediction of High Mitotic Count in Gastrointestinal Stromal Tumors Using CT Features and Serologic Indicators: An Interpretable Model with Multicenter External Validation

Background: Accurate preoperative identification of high mitotic count is important for risk stratification in gastrointestinal stromal tumors (GISTs), yet biopsy is invasive and conventional imaging alone does not directly reflect microscopic proliferative activity. We aimed to develop and externally validate an interpretable model integrating CT features and serologic indicators. Methods: This multicenter retrospective study included 802 patients from three Shanxi hospitals, randomly split into a training cohort (n = 562) and an internal validation cohort (n = 240), plus an external validation cohort from Dalian (n = 255). LASSO and multivariable logistic regression were used for feature selection and nomogram construction. Five individual machine-learning models and 31 stacking ensembles were trained, and SHAP was used to interpret model behavior. Results: Tumor size, liquefaction/necrosis, coarse vessel sign, peritumoral fat stranding, platelet-to-lymphocyte ratio, and albumin-to-fibrinogen ratio were independent predictors of high mitotic count. Among individual models, SVM achieved the highest internal-validation AUC (0.866). The best stacking model (SVM+ANN+Logit) reached an AUC of 0.867 on internal validation and 0.955 on external validation, with good calibration and favorable decision-curve performance. The gain over the best single model was small but consistent. Conclusion: An interpretable model combining CT and serologic features may provide a practical non-invasive tool for preoperative estimation of mitotic count in GISTs. Prospective validation is still needed before routine clinical implementation. The flowchart illustrates the process of patient data collection and model development for a study on gastrointestinal stromal tumors. Patients were collected from three hospitals between January 2017 and January 2025. After screening, 802 patients were included and divided into a training set of 562 patients and an internal validation set of 240 patients, with subgroups defined by mitotic index > 5/50 HPFs and ≤ 5/50 HPFs. An external test set from another hospital included 255 patients. The workflow shows variable selection using LASSO and multivariable logistic regression, followed by the development of a logistic regression-based nomogram model and several machine learning models, including XGBoost, LightGBM, Random Forest, SVM, and ANN. Model performance was compared, and a stacking ensemble model was established. SHAP analysis was also included for model interpretation.GIST study workflow: data collection, model development, performance comparison, and SHAP interpretation. Keywords: gastrointestinal stromal tumors, mitotic count, stacking ensemble learning, computed tomography, serological indicators, risk stratification