MULTIMODAL FOUNDATION MODELS FOR PREDICTIVE ONCOLOGY: UNIFYING RADIOMICS, PATHOMICS, GENOMICS, AND ELECTRONIC HEALTH RECORDS

Abstract

Predictive oncology is undergoing a paradigm shift through the emergence of multimodal foundation models capable of integrating radiomics, pathomics, genomics, electronic health records (EHRs), and other heterogeneous biomedical data into unified computational frameworks for precision cancer care. Conventional artificial intelligence (AI) models have achieved remarkable success in individual diagnostic or prognostic tasks; however, their dependence on task-specific training and isolated data modalities limits their ability to fully characterize the biological complexity of cancer. Recent advances in foundation AI models, multimodal transformer architectures, graph neural networks, self-supervised learning, and generative artificial intelligence have enabled generalized representation learning across radiological imaging, digital pathology, genomics, transcriptomics, proteomics, metabolomics, epigenomics, spatial biology, laboratory biomarkers, liquid biopsy, wearable physiological monitoring, electronic health records, and longitudinal clinical outcomes. These intelligent computational systems support precision diagnosis, biomarker discovery, molecular characterization, prognostic prediction, therapeutic optimization, immunotherapy selection, adaptive disease monitoring, and personalized clinical decision support. Emerging technologies including multimodal large language models, federated learning, reinforcement learning, retrieval-augmented generation, explainable artificial intelligence, and agentic AI further strengthen predictive oncology by enabling collaborative, privacy-preserving, and continuously adaptive biomedical intelligence. Despite substantial technological progress, important scientific, technical, ethical, and regulatory challenges remain regarding multimodal data harmonization, computational scalability, interpretability, interoperability, cybersecurity, clinical validation, and equitable implementation. This review provides a comprehensive overview of multimodal foundation models for predictive oncology, emphasizing the integration of radiomics, pathomics, genomics, and electronic health records as a transformative framework for personalized cancer medicine.