Bending the Learning Curve for EHR Research via Knowledge-Driven Online Multimodal Automated Phenotyping System

Electronic health records (EHRs) hold great promise for translational research but remain difficult to use at scale because diagnostic codes are noisy, disease-relevant features are hard to identify, high-quality labels are limited and patient-level data sharing is often restricted. We introduce Knowledge-driven Online Multimodal Automated Phenotyping (KOMAP), a unified system for automated, privacy-preserving EHR phenotyping using summary statistics. The system is organized around three original and integrated layers. First, the Multi-source Representation Learning (MultiReL) embedding layer fuses EHR co-occurrence information from multiple health systems with biomedical language-model representations into a shared semantic space. Second, the Online Narrative and Codified feature Engine (ONCE) retrieves and ranks disease-relevant codified and narrative concepts using MultiReL and local EHR support. Third, the KOMAP phenotyping layer trains, validates, and transports disease-specific algorithms from simple summary statistics rather than individual-level patient records. Validation in four healthcare centers demonstrated the ability of KOMAP to support patient subtyping, generate highly accurate phenotyping algorithms, and improve statistical power in downstream phenome-wide association studies (PheWAS). In a PheWAS of a statin-related genetic variant, KOMAP identified significant protective associations for alopecia and seborrheic dermatitis that were not found by standard methods. By combining knowledge-guided feature retrieval with summary-statistics-based phenotyping, the proposed framework reduces privacy, computational, and technical barriers to high-throughput EHR research and supports scalable multi-institutional biomedical discovery.