Towards Precise Treatment Response and Overall Survival Prediction in Gastric Cancer via Interpretable and Robust Multimodal Data Integration

Abstract Gastric cancer (GC) emerges as a significant global malignancy, ranking as the fifth most prevalent cancer in 2022. The integration of multimodal data provides a more holistic understanding of the disease, enhancing predictions of treatment response and overall survival for GC patients. Nevertheless, the inherent heterogeneity and potential incompleteness of multimodal data pose formidable challenges for effective integration and in-depth analysis. Furthermore, the lack of interpretability in the prediction models hinders their trustworthiness and acceptance in clinical practice. Therefore, this study presented an interpretable and robust multimodal data integration framework for GC analysis (iMD4GC), facilitating effective integration of diverse data sources to predict treatment response and overall survival. Three multimodal datasets were collected for model evaluation: GastricRes (n=698) for response prediction, GastricSur (n=801) for survival analysis, and TCGA-STAD (n=400) for further survival analysis. The iMD4GC achieved an area under the receiver operating characteristic curve (AUC) of 0.803 (95% CI 0.680-0.925) on GastricRes, a concordance index (c-index) of 0.715 (95% CI 0.646-0.784) and a time-dependent AUC of 0.739 (95% CI 0.644-0.833) on GastricSur, and a cindex of 0.661 (95% CI 0.504-0.818) and a time-dependent AUC of 0.690 (95% CI 0.500-0.881) on TCGA-STAD. The comparative analysis underscores the reliability of iMD4GC in integrating diverse multimodal data and addressing missing modalities. Further interpretability analysis identified pivotal factors across various data modalities, validating the alignment with clinical expertise and providing essential insights for informed decision-making in clinical contexts. The robustness and scalability hold significant implications for clinical practice, enabling precision oncology through artificial intelligence and multimodal data integration.