Shortening the diagnostic odyssey: multimodal patient similarity networks and explainable artificial intelligence for rare disease diagnosis (360G-Wellcome-320092_Z_24_Z)

While individually uncommon, rare diseases (RDs) affect around 300 million people worldwide. Patients with RDs are faced with a diagnostic odyssey despite advances in genomic testing. Even when an accurate diagnosis is obtained, the majority of RDs lack approved treatments. RD research is faced with challenges including limited data, difficulties in interpreting the data, and disease heterogeneity. Without effective tools and methodologies to address these challenges, patients with RDs will continue to face significant barriers to accurate diagnosis, timely intervention, and access to specialized care. We propose to develop a suite of models that integrate UK and worldwide (DECIPHER, UK Biobank, Genomics England) multimodal data at the patient level in patient similarity networks (PSNs) to improve characterization and diagnosis of RDs. We also aim to develop graph neural networks (GNNs) to identify undiagnosed patients in paediatric and adult populations, with a focus on interpretability and explainability for enhanced decision-making support. This is a clinically intuitive approach, since it mimics how clinicians think: “Based on the patients I have seen and treated so far, which group does this patient belong to?” This approach can improve our understanding and diagnosis of RDs, ultimately leading to better patient outcomes and advancing medical research.