Optimizing Real-Time Phenotyping in Critical Care Using Machine Learning on Electronic Health Records

Objective: Real-time understanding of a patient’s physiological state is essential in critical care, yet machine learning methods for real-time phenotyping remain underdeveloped due to signal complexity and lack of timestamped diagnoses in Electronic Health Records (EHRs). We address this by introducing a method that continuously updates phenotype predictions from evolving biomedical signals, bridging signal processing and machine learning to enable continuous clinical decision support. Methods: We developed a modular real-time phenotyping system that efficiently handle variable-length, irregularly sampled EHR data with missing values. A multi-task Long Short-Term Memory (LSTM) model jointly learns 561 diagnostic labels, capturing long-range nonlinear dynamics while providing stable objective estimation and robust handling of class imbalance and low-prevalence phenotypes. Results: Our method achieved good (AUC-ROC  ≥ 0.8) or excellent (AUC-ROC  ≥ 0.9) diagnostic performance for 69% and 30% of 561 phenotypes, respectively. Notably, 42% and 7% reached these thresholds using only the first recorded signals, supporting the feasibility and clinical potential of early phenotyping. Using inferred phenotypes as inputs for downstream outcome tasks (in-hospital, 24-hour, ICU, and 30-day mortality) enabled evaluation against ground-truth labels while extending traditional clinical scores to continuous prediction and improving interpretability through risk attribution to clinically meaningful latent states. Conclusion: The proposed framework offers a promising approach for accurate, real-time EHR-based phenotyping across a wide range of conditions by continuously refining predictions from heterogeneous biomedical signals that reflect non-stationary disease dynamics. Comparable performance across acute and chronic phenotypes supports its ability to model underlying disease mechanisms, while interpretability analysis confirms its alignment with the data-generating process and sensitivity to healthcare-related biases.