In this work, we investigate the feasibility of prospectively identifying patients at risk for rare diseases using longitudinal electronic health record (EHR) data. Leveraging a large real-world cohort of over 40,000 patients, including 879 patients identified through a curated rare disease dataset, we construct a temporally grounded prediction framework that estimates risk at multiple time horizons up to one year prior to the first documented rare disease event. We developed deep learning models designed to capture both intra-visit clinical heterogeneity and inter-visit temporal progression, enabling the detection of subtle and distributed early signals across patient histories. Our evaluation demonstrates that meaningful predictive performance can be achieved even in a moderately imbalanced setting (~2% prevalence), with improved signal strength observed at longer prediction horizons. This work could support earlier referral for genetic testing and reduce the years-long diagnostic delays that rare disease patients too often face.

TBA