Mamba-based AI Improves Patient Subtyping from EHR Data

Subtyping patients based on their electronic health record (EHR) data is critical for advancing precision medicine, but the inherent complexity and irregularity of temporal EHR datasets pose significant challenges. This research introduces a novel approach to overcome these difficulties. The study proposes a self-supervised model built on a Mamba-based architecture, designed to learn robust and effective representations from longitudinal EHR data. This representation learning capability is then applied to improve patient subtyping. Extensive evaluations on both public and private real-world EHR datasets show that the Mamba-based model significantly outperforms competitive baseline models. The findings offer valuable insights for classifying and clustering patients based on their temporal health records, paving the way for more personalized medical interventions.