Abstract
Diverse clinical data, encompassing electronic health records, genomic profiles, and medical imaging, encode complex relationships between clinical variables, patients, and outcomes. Graph-based learning provides a natural framework for modeling these relationships, but effective representation learning in the face of data heterogeneity and the need of reliable uncertainty estimates for high-stakes clinical predictions remain open challenges. This dissertation develops a novel machine learning framework that integrates graph neural networks for relational modeling, multimodal learning for comprehensive data fusion, and Gaussian process-based uncertainty estimation for reliable predictions across two critical clinical tasks: sepsis onset prediction and breast cancer molecular subtype classification. This framework employs multimodal learning strategies to fuse information beyond structured EHRs, including temporal and relational dependencies from the MIMIC-IV dataset for sepsis and clinical, multi-omics, and imaging data for breast cancer. Spectral-normalized graph transformer convolution layers extract structured representations, while a random feature-based Gaussian process layer captures and quantifies predictive uncertainty. Integrated gradients provide interpretable attributions across temporal features, offering insight into global and patient-level prediction drivers. Uncertainty-aware modeling enhances reliability, providing well-calibrated confidence estimates that support clinical decision-making. This research demonstrates that the novel combination of graph-based learning, multimodal data fusion, and uncertainty estimation significantly improves predictive modeling in healthcare by addressing key challenges in data integration, representation learning, and uncertainty quantification. Future work will explore scaling multimodal integration, improving real-time inference, and extending interpretability methods to support clinician-centered deployment.