S1302 Advancing Gastrointestinal Disease Diagnosis and Prognosis With GastroEndo-Seq: An AI-Driven Approach for Stage Classification and Disease Progression Prediction

Introduction: Gastrointestinal disorders pose a substantial healthcare burden given their prevalence and risk of malignancy or serious complications. Early detection and accurate staging are vital to improve outcomes and reduce costs, yet current diagnostics depend largely on physician expertise with limited objective tools for forecasting progression. GastroEndo-Seq addresses this gap by combining artificial intelligence and endoscopic imaging to predict GI disease stages and simulate clinical trajectories using the multi-center GastroVision dataset, aiming to enhance diagnostic accuracy and support personalized treatment. Methods: GastroEndo-Seq was trained on 8000 expert-annotated endoscopic images from the GastroVision repository (Baerum Hospital, Norway and Karolinska University Hospital, Sweden), encompassing 4 stages: Normal, Inflammatory, Pre-Malignant, and Advanced/Malignant. Annotations were confirmed by experienced GI endoscopists. We compared 4 architectures—Vision Transformer (ViT), DenseNet-121, ResNet-50, and a bespoke convolutional neural network—and selected ViT for its superior global feature extraction. To bolster performance in underrepresented classes (<100 images), we applied few-shot learning. Model interpretability was enhanced using Score-CAM heatmaps to highlight critical image regions. Performance was evaluated via accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Results: GastroEndo-Seq achieved 80% overall accuracy across the 4 stages. Class-specific AUCs ranged from 0.80 to 1.00, with a micro-average AUC of 0.98, demonstrating robust multi-class discrimination. Few-shot learning yielded 81% accuracy in sparsely populated categories, effectively mitigating data-scarcity issues. Score-CAM correctly identified key regions of interest in 93% of cases, confirming model transparency and clinical relevance. Conclusion: GastroEndo-Seq reliably classifies GI disease stages and simulates progression trajectories that align with clinical observations. The integration of few-shot learning addresses data imbalance, while Score-CAM visualizations improve interpretability. This approach holds promise to reduce diagnostic variability and inform personalized management strategies. Future work will expand the dataset, incorporate additional imaging modalities, and validate real-time deployment in clinical settings.