RadRelBench: A Benchmark for Deep Learning on Multimodal Relational Databases of Chest Radiographs

Traditional Machine Learning approaches applied to Electronic Health Records (EHRs) often require flattening complex relational databases into single tabular matrices. This conventional preprocessing step leads to the destruction of the sequential clinical hierarchy, the creation of highly sparse feature spaces, and a significantly increased risk of temporal data leakage. To overcome these fundamental limitations, this thesis proposes and validates a Multimodal Relational Deep Learning pipeline leveraging the open-source RelBench framework. By natively mapping the relational schema of the MIMIC-IV database into a Heterogeneous Temporal Graph, the proposed architecture preserves the intrinsic chronological order of patient admissions, laboratory results, and clinical interventions. Furthermore, this methodology successfully extends the relational structure to encompass unstructured visual data by seamlessly integrating high-dimensional Med-SigLIP embeddings, extracted from MIMIC- CXR-JPG chest radiographs, directly into the graph topology. The predictive capabilities of the architecture are empirically evaluated across three distinct clinical tasks: ICU Length of Stay (administrative), In-Hospital Mortality (prognostic), and CXR Multilabel Pathology Classification (diagnostic). The core learning mechanism, utilizing HeteroGraphSAGE, effectively aggregates multi-hop clinical and visual context through strict temporal neighborhood sampling, rigorously guaranteeing the absence of future data leakage. Experimental results demonstrate the clear superiority of the Multimodal Graph Neural Network. The proposed model consistently outperforms both traditional feature-engineered baselines (LightGBM) and unimodal network variants (Tabular-Only and Image-Only GNNs). Ultimately, this work establishes a robust, leakage-free, and highly scalable paradigm for transforming raw, heterogeneous hospital databases into accurate predictive intelligence.