S480 Advanced Machine Learning Framework for Predicting Polyp Evolution and Malignancy Risk in Gastroenterology: A Novel Approach to Personalized Surveillance and Early Intervention

Introduction: Early prediction of polyp progression and malignancy is critical in gastroenterology to guide timely intervention for high-risk cases. Conventional approaches often overlook the dynamic nature of lesion development and fail to integrate multimodal data, limiting their prognostic power. To address these gaps, we present a novel machine-learning framework that combines longitudinal simulation, advanced imaging analysis, and patient metadata to enhance prediction of polyp evolution and risk stratification. Methods: We used the ERC PMP-v5 dataset, comprising 796 high-resolution images and 21 annotated endoscopic videos from 217 patients, to train and test our models. Two complementary pipelines were developed. The polyp evolution model: We generated synthetic time-series sequences over 30 months at 6-month intervals to simulate growth trajectories. A Long Short-Term Memory (LSTM)-based regressor, trained on these artificial sequences, predicted size changes with a mean absolute error (MAE) of 0.0689. Concurrently, a transformer-based classifier distinguished morphological transitions (e.g., tubular to villous or adenocarcinoma) with 85 percent accuracy. These outputs yield interval-specific clinical insights. The second pipeline, fusion model for malignancy and recurrence, used extracted imaging features via a Vision Transformer and encoded clinical metadata (age, family history, histology, etc.). A fusion network combined these inputs to predict malignancy risk and recurrence, achieving 99.89 percent accuracy. Patients were stratified into low, moderate, and high-risk categories to inform surveillance intervals. Results: The LSTM regressor reliably forecasted critical size milestones—predicting a 0.09 cm tubular polyp at 6 months (low risk) and a 0.30 cm villous lesion at 30 months (high risk). The transformer classifier accurately identified high-risk morphology changes, enabling early clinical action. The fusion model’s near-perfect accuracy in malignancy and recurrence prediction facilitated clear risk stratification: low-risk patients were advised follow-up in 1–3 years, moderate-risk in 6–12 months, and high-risk within 3–6 months. Conclusion: By integrating LSTM and transformer architectures with a ViT-based fusion approach, our framework delivers precise predictions of polyp growth, morphological transformation, and malignancy risk. This method supports personalized surveillance recommendations and prioritizes high-risk cases, underscoring the potential of advanced machine learning to transform decision-making in gastroenterology.