A Bayesian Nonparametric Model with an Informative Visit Process: Using Electronic Health Records to Evaluate Body Mass Index for Diabetes Screening
Conference
65th ISI World Statistics Congress 2025
Format: IPS Abstract - WSC 2025
Keywords: electronic-health-records
Session: IPS 806 - Advances in Handling Missing Data for EHR and Causal Inference
Monday 6 October 2 p.m. - 3:40 p.m. (Europe/Amsterdam)
Abstract
Type 2 diabetes screening among Asian Americans faces a public health challenge. Although a body mass index (BMI) of 23 kg/m2 or higher has been recently recommended to refer Asians to diabetes diagnostic testing, cross-sectional research has suggested the poor predictive performance of the new cutoff. The findings, however, are inconclusive because a cross-sectional design cannot guarantee that BMI was measured before diabetes onset. More generally, the utility of BMI itself for diabetes screening is uncertain. Using longitudinal electronic health records from a New York City hospital system, we evaluate the effectiveness of alternative static BMI cutoffs versus patient BMI trajectories for diabetes screening in Asians. We build a Bayesian nonparametric joint model of longitudinal BMI and time-to-diabetes diagnosis. To account for an informative visit process whereby a patient’s visit process may be associated with underlying health status, we add a recurrent event submodel for gap times between a patient’s clinic visits. We apply the Bayesian model to estimate measures of predictive performance over follow-up. We conduct a simulation study to examine the statistical properties of the proposed method.