(B-042) Validation of a claims-based algorithm for body mass index (BMI) and assessment of its utility to balance BMI in active comparator new user studies

Background: Uncontrolled confounding by unmeasured clinical measures is a common concern in comparative effectiveness studies of anti-hyperglycemic medications using administrative claims data in the United States. Claims-derived predicted values of these unmeasured clinical factors provide one potential approach to address this limitation.

Objectives: To: (1) Externally validate a claims-based prediction model for BMI in Medicare fee-for-service (FFS) beneficiaries using linked electronic health record (EHR) data and (2) assess the utility of this predicted BMI to reduce imbalance in measured BMI in the setting of an active comparator new user study in this linked dataset.

Methods: We linked patients aged ≥ 65 with at least one encounter with a large, statewide health system in North Carolina (USA) to Medicare FFS claims during 2014-2019. We externally validated a claims-based prediction model for BMI developed in Medicare, against EHR recorded BMI using claims in a 181-day period either centered at (overlap window) or prior to (lookback window) the measured BMI. Model performance was assessed with plots and summary metrics (mean error, slope, R2 etc.). Utility of these predicted values, anchored in the lookback window, to account for baseline imbalance in BMI, was assessed using standardized mean differences (SMD) in distributions of most recent measured BMI in the 6-months prior to drug initiation. Glucagon-like peptide 1 receptor agonist (GLP-1) initiators were compared to one of 4 potential active comparators before and after accounting for claims-predicted BMI using inverse probability of treatment weights.

Results: Compared to measured BMI, the claims-based predictive model performed similarly using either the overlap (N=26,314) or lookback (N=28,624) windows. Calibration plots suggested poor calibration in BMI ranges < 27 and >36 (absolute mean error > 2.5 kilograms/meter2). Measured BMI was highest among new users of GLP1. SMDs for baseline measured BMI for GLP1 were: 0.26 vs. SGLT2i (GLP1 N=554, SGLT2i N=315), 0.44 vs. DPP4i (GLP1 N=490, DPP4i N=689), 0.48 vs. SU (GLP1 N=373, SU N=904), and 0.56 vs. LAI (GLP1 N=362, SGLT2i N=781). After weighting to account for claims-predicted BMI, imbalance in measured BMI improved across all comparator groups to < 0.2.

Conclusions: A claims-based prediction model for BMI showed poor calibration for measured BMI in this external validation study. While accounting for the claims predicted BMI values did improve baseline imbalance in measured BMI, the choice of active comparator remains a key study design strategy for reducing baseline differences in BMI, if it were measured, in comparative effectiveness studies of anti-hyperglycemic agents.