Machine learning-based prediction of long-term new-onset diabetes mellitus risk after pancreaticoduodenectomy using radiomics

Background Pancreaticoduodenectomy carries substantial metabolic consequences, with 20–32% of patients developing new-onset diabetes mellitus (NODM) within three years, leading to increased morbidity and healthcare burden. Current predictive models relying primarily on clinical variables demonstrate limited accuracy, underutilizing tissue-level information available in routine CT imaging. This study aimed to develop and validate a multimodal machine learning framework integrating clinical data with CT-derived radiomics features for long-term NODM risk prediction. Methods This retrospective cohort study analyzed 126 patients who underwent pancreaticoduodenectomy at Gachon University Gil Medical Center (2005–2023). Using PyRadiomics, 186 radiomic features were extracted from preoperative and postoperative CT scans (93 features per timepoint). Combined with 10 clinical variables (196 total features), Recursive Feature Elimination identified 10 key predictors. Logistic Regression, Support Vector Machine, Random Forest, and Extreme Gradient Boosting were evaluated using 5-fold cross-validation. SHAP analysis ensured model interpretability. Results Long-term NODM developed in 47 patients (37.3%). The Logistic Regression model demonstrated optimal performance with AUC 0.77 (95% CI: 0.68–0.84), sensitivity 70% (95% CI: 0.57–0.83), and specificity 72% (95% CI: 0.62–0.82). Key predictors included pancreatic volume changes, preoperative hypertension, and texture features (Strength, GrayLevelNonUniformity) from both imaging timepoints. The multimodal approach significantly outperformed clinical-only models ( P < .05). Subgroup analyses confirmed consistent model performance across gender ( P = .72) and age groups ( P = .83). Conclusions The proposed approach that integrates CT radiomics with clinical data quantitatively improved the prediction performance for NODM after pancreatectomy. This multimodal strategy offers a more robust alternative to single-modality models and may facilitate personalized risk stratification and targeted postoperative surveillance.