A dynamic graph-based PU-learning approach for sequential recommendation in rare disease detection

Introduction: Rare diseases affect over 300 million people globally; their low prevalence often leads to high misdiagnosis rates. While Electronic Health Record (EHR) data offers opportunities for early detection, existing predictive models struggle to capture the complex, time-sensitive sequence of patient journeys. Materials and methods: To address this, we propose a Dynamic Graph-based Positive-Unlabeled Learning approach (DPU-SR) to forecast patient trajectories and identify undocumented rare diseases. Our framework reconstructs patients and clinical events into dynamic subgraphs, utilizing time-aware information propagation to evaluate how sequential clinical actions influence health outcomes. To address EHR limitations where records indicate diagnosed conditions but cannot definitively rule out others, we implement dynamic structure-aware negative sampling within a Positive-Unlabeled (PU) learning setting. Results: After benchmark validation, we deployed the DPU-SR algorithm on a real-world EHR dataset of approximately 1.5 million candidates to detect unrecognized multiple myeloma (MM). By capturing the temporal evolution of clinical interactions, our approach achieved an average 6.3% precision among the top 1000 scored patients from each testing cohort, showing a 20% improvement over the commonly used production models in the healthcare industry. Conclusions: Transforming sequential EHR data into dynamic graphs provides an effective, scalable artificial intelligence tool for population-wide disease surveillance. Enhancing early detection precision provides significant predictive enrichment and the potential to inform early intervention, offering a scalable foundation to support public health practice, optimize resource allocation, and proactively manage patient care.