Abstract 4364499: Performance of the Diabetes Outcomes Model in the United States (DOMUS) in a Large, Contemporary US Claims Dataset

Abstract

Objectives: The Diabetes Outcomes Model in the US (DOMUS) was developed using 13 years (2005–2017) of data from Kaiser Permanente Northern California to project health outcomes and inform economic evaluations for patients with type 2 diabetes (T2D). However, its performance in broader, more diverse populations remains unclear. This study evaluated DOMUS using Optum Market Clarity, a large claims–electronic health record (EHR) linked dataset, following individuals with T2D in the US from 2008 to 2023. Methods: DOMUS was used to predict the 10-year cumulative incidence of macrovascular (atrial fibrillation, angina, congestive heart failure, myocardial infarction [MI], ischemic heart disease [IHD], peripheral vascular disease [PVD], and stroke) and microvascular (retinopathy, end-stage renal disease [ESRD], foot ulcer, and blindness) complications in over 70,000 participants. Model calibration (comparison of predicted vs. observed cumulative incidence) and discrimination (c-statistics) were assessed for each endpoint. For poorly calibrated outcomes, recalibration was performed by updating the model constant using a grid search. Results: Before model recalibration, DOMUS underpredicted most macrovascular complications over 10 years (absolute mean difference: 1%–9%), except for PVD (overpredicted by 6%) and stroke (accurately predicted). Among microvascular outcomes, proliferative retinopathy and ESRD were well predicted, while foot ulcer and blindness were underestimated, and non-proliferative retinopathy was overestimated. Nonetheless, DOMUS demonstrated strong discrimination (c-statistics >0.75 for all endpoints). After recalibration, the predicted 10-year cumulative incidences for all outcomes were within 1% of observed rates. Conclusions: Original DOMUS demonstrated strong discrimination across all outcomes though did not predict many outcomes well in Optum Market Clarity dataset. However, after recalibration, the model achieved accurate predictions within 1% of the observed cumulative incidence at 10 years for all outcomes. These findings highlight the importance of validating and refining risk prediction models when applied to diverse, real-world populations.