Data-driven discovery of medication effects on blood glucose from electronic health records.

Abstract

Blood glucose (BG) in hospitalized patients is influenced by numerous clinical factors, including medications not traditionally associated with glycemic control. To better characterize these effects, we analyzed electronic health record data from 97,281 inpatient encounters (2014-2022), capturing 3,009,686 point-of-care BG measurements. We extracted over 300 variables-medications, labs, and socio-demographics-and used Lasso, ridge, and elastic net regression for predictive modeling, alongside propensity score matching (PSM) for causal inference. While Lasso reduced multicollinearity, it often assigned implausible coefficient directions. In contrast, PSM yielded clinically consistent and interpretable estimates, identifying 55 variables significantly associated with BG changes, without shrinking coefficients to zero of known BG-modulating drugs. Findings were validated in a 2022-2024 test set of 27,847 encounters. This work highlights the value of causal inference in observational EHR analysis and identifies both established and under-recognized (e.g., cholecalciferol) medication effects on BG, offering insights that inform safer inpatient glycemic management.