S688 Integrated Clinical and Endoscopic Imaging Model Accurately Predicts Colorectal Polyp Recurrence

Introduction: Identifying patients at high risk for colorectal polyp recurrence remains an unmet need in endoscopic practice. Conventional guidelines rely on polyp size and histologic grade but often misclassify borderline cases. By blending endoscopic image analysis with structured clinical data, a predictive model may provide more accurate, personalized surveillance recommendations and reduce unnecessary procedures. Methods: We retrospectively assembled a cohort from the Endoscopic Image and Video Dataset for Recognition of Colorectal Polyps Morphology and Pathology-v5 repository, pairing high-resolution polyp images with corresponding clinical metadata: size, morphology, dysplasia grade, and prior recurrence. We standardized clinical inputs and resize all images to 224 × 224 pixels. An EfficientNet-V2 backbone extracted radiomic features, while a fully connected network processed clinical variables. We fused these embeddings in a late-fusion stage, then applied an ensemble prediction head combining Gradient Boosting and a multilayer perceptron. We trained the model using binary cross-entropy loss, calibrated probabilities via isotonic regression, and stratified outputs into Low (36-month follow-up), Moderate (12–24 months), and High (3–6 months) risk categories. Model performance was assessed on a held-out test set using accuracy, AUC, sensitivity, specificity, and test loss. SHAP analysis determined feature contributions. Results: On unseen data, the fusion ensemble achieved 91.2% accuracy and an AUC of 0.932. Sensitivity measured 89.7%, specificity reached 93.5%, and test loss was 0.0013. Calibration curves showed minimal deviation between predicted and observed recurrence probabilities (Brier score 0.038). SHAP rankings consistently placed polyp size first, followed by dysplasia grade and recurrence history; image-derived textural descriptors improved discrimination in equivocal cases. Risk stratification aligned with actual recurrence rates: Low—5.4%, Moderate—23.8%, High—68.7%. Feature importance remained stable across cross-validation folds. Conclusion: By integrating endoscopic radiomics with clinical parameters, our model delivered high accuracy and reliable risk stratification for polyp recurrence. Key predictors—lesion size, histologic grade, and recurrence history—align with established risk factors, while radiomic features capture subtle image patterns. This approach offers a scalable tool to tailor surveillance intervals and may reduce unnecessary colonoscopies. Prospective validation will determine its generalizability across diverse practice settings.