Explainable Machine Learning for Early Prediction of Surgical Necessity in Gastrointestinal Emergencies: A Multimodal Diagnostic Study

Background/Objectives: Acute gastrointestinal (GI) emergencies require timely and accurate prediction of surgical necessity to avoid delayed intervention and improve patient outcomes. Traditional scoring systems offer limited accuracy and fail to integrate multimodal data. This study aimed to develop and validate an explainable machine learning model for early prediction of surgical necessity in patients presenting with GI emergencies. Methods: A retrospective cohort of 1032 consecutive adult patients admitted with acute GI emergencies at a tertiary referral center (2019–2024) was analyzed. Three predictive models were developed: logistic regression, Random Forest, and XGBoost. Features included clinical, laboratory, and contrast-enhanced CT imaging variables available within the first 24 h. Model performance was evaluated using AUC, sensitivity, specificity, PPV, NPV, and F1-score. Shapley Additive Explanations (SHAP) were applied for global and individual-level interpretability. The study followed STROBE and TRIPOD+AI reporting guidelines. Results: Surgical intervention was required in 312 patients (30.2%). The XGBoost model achieved the highest AUC (0.89; 95% CI: 0.86–0.92), outperforming Random Forest (AUC 0.86) and logistic regression (AUC 0.79), with sensitivity 0.84, specificity 0.81, and NPV 0.90. The most influential predictors were serum lactate, CT findings (free intraperitoneal air, bowel ischemia), IL-6, and shock index. Decision curve analysis confirmed net clinical benefit across threshold probabilities of 10–70%. Subgroup performance remained robust across diagnostic categories (AUC 0.87–0.91). Conclusions: An explainable XGBoost model integrating early clinical, laboratory, and imaging data accurately predicts surgical necessity in GI emergencies and outperforms traditional scoring systems. SHAP-based explainability supports clinical adoption and transparency. Prospective multicenter validation is warranted. The positive predictive value of 0.74 indicates that approximately one in four patients flagged as requiring surgery may not need operative intervention. The model should be regarded as a decision-support adjunct, rather than a standalone surgical decision tool, that is most relevant in settings where immediate experienced surgical judgment is limited.