Binary PheNorm enhances EHR phenotyping with binary labels, boosting AUC up to 12%

Electronic health record (EHR) studies often lack gold-standard phenotype labels because manual chart review is time-consuming and expensive. Weakly supervised approaches, like PheNorm, use silver-standard labels—such as diagnosis code counts, NLP mentions, or lab thresholds—to train phenotype classifiers. However, the original PheNorm was designed for count-valued silver labels, requiring log transformation and Gaussian mixture modeling, which are ill-suited for binary indicators (e.g., whether a medication was mentioned or a lab test exceeded a threshold). This limits its utility in many clinical settings where only presence/absence signals are available.

To address this gap, the team developed Binary PheNorm, which adapts the corruption‑and‑regression denoising step to accept binary silver labels directly. It produces a continuous phenotype score without the need for EM calibration, and includes a lasso‑regularized variant for high‑dimensional settings with many candidate features. In simulated and real‑world datasets, Binary PheNorm using a single binary label (epinephrine mention for anaphylaxis, lipase threshold for acute pancreatitis) achieved AUCs of 0.891–0.892 and 0.805–0.819 respectively—substantial improvements over raw indicator baselines (0.793, 0.736). Combining binary and count labels often yielded further gains. The method offers a practical, scalable tool for automated phenotyping in EHR research.