Automating psychiatric distress screening in ophthalmology using electronic health records: A case study in scalable computational algorithms for Bayesian inference
Samuel I. Berchuck, PhD Senior Research Associate Department of Statistical Science Duke University
Automating psychiatric distress screening in ophthalmology using electronic health records: A case study in scalable computational algorithms for Bayesian inference
-Abstract: In patients with ophthalmic disorders, psychosocial risk factors play an important role in morbidity and mortality. Proper and early psychiatric screening can result in prompt intervention and mitigate its impact. Because screening is resource intensive, we developed a framework for automating screening using electronic health record (EHR) data. EHR data in our study came from the Duke Ophthalmic Registry, a retrospective EHR database that contains medical and clinical records for all patients seen at the Duke University Eye Center since 1993. In order to account for the high-dimensional and longitudinal nature of the modeling task, we developed a scalable version of the generalized linear mixed model (GLMM). Standard posterior sampling algorithms, such as Markov chain Monte Carlo (MCMC) procedures, are not inherently scalable and have limited utility in large datasets. To overcome this limitation, we introduced a stochastic gradient MCMC (SGMCMC) algorithm that uses mini-batch samples to approximate the true gradient over the whole dataset. As part of the algorithm, we developed a Monte Carlo estimator to approximate the gradient of the intractable marginal log likelihood associated with GLMM. Through simulation, we show that our method scales to large data settings while maintaining proper uncertainty quantification. Finally, we demonstrate the utility of the scalable GLMM to classify psychiatric distress for ophthalmology clinic encounters through comparison with existing machine learning prediction models.
Virtual format: Zoom link to follow 25 January 2023 1:30pm