S685 Multimodal Integration of Histopathology and Clinical Data Predicts Survival in Colorectal Adenocarcinoma

Introduction: Accurate prognostication in colorectal adenocarcinoma requires combining cellular morphology with patient factors. Histopathology may miss systemic influences, and clinical variables lack spatial context. Integrating whole‐slide image features with clinical metadata could yield precise, interpretable survival models. Methods: We used 428 TCGA-COAD microsatellite instability (MSI)/microsatellite stability (MSS) cases with whole-slide images and clinical data. Slides were tiled into 224 × 224-pixel patches. An EfficientNet-V2 backbone generated 1,280-dimensional feature vectors. Attention-based multiple instance learning aggregated patch embeddings into slide-level representations. Clinical variables (stage, MSI/MSS status, age, Charlson index, lymphovascular invasion) were encoded via a 3-layer network into a 256-dimensional embedding. We concatenated embeddings and trained a survival regression network using a Cox proportional hazards loss. Data were split into training (70%), validation (10%), and test (20%) sets stratified by stage and MSI status. We optimized with Adam (learning rate = 1 × 10⁻⁴), used early stopping on validation C-index, and calibrated risk scores. We evaluated on the test set using Harrell’s C-index, concordance accuracy, weighted F1-score, and Brier score. Kaplan–Meier curves and log-rank tests assessed risk tertiles. SHAP quantified feature importance and attention heatmaps localized critical histologic regions. Results: On the test set, the fused model achieved C-index = 0.873 and concordance accuracy = 86.4%. The weighted F1-score was 87.1%, and the Brier score was 0.142. Kaplan–Meier curves showed median survival of 18 months for high-risk versus 76 months for low-risk (log-rank P < 0.001). Attention heatmaps highlighted tumor margins and peritumoral stroma. SHAP analysis identified MSI/MSS status and tumor stage as top clinical predictors: microsatellite stability and advanced stage increased risk. Histologic features of dense immune infiltration and well-differentiated glands lowered risk, while glandular disarray and desmoplastic stroma raised risk. These rankings remained consistent across ten cross-validation folds. Conclusion: Integrating spatial histopathology with clinical variables yields accurate, interpretable survival predictions in colorectal adenocarcinoma. Key predictors align with known risk factors, supporting clinical utility. Prospective validation will assess generalizability and guide integration into routine workflows.