Estimating the prevalence of diabetic retinopathy in electronic health records with massive missing labels

Objective: The paper aims to address the problem of massive unlabeled patients in electronic health records (EHR) who potentially have undiagnosed diabetic retinopathy (DR). It is desired to estimate the actual DR prevalence in EHR with 96 % missing labels. Materials and methods: The Cerner Health Facts data are used in the study, with 3749 labeled DR patients and 97,876 unlabeled diabetic patients. This extensive dataset spans the demographics of the United States over the past two decades. We implemented state-of-art positive-unlabeled learning methods, including ensemble-based support vector machine, ensemble-based random forest, and Bayesian finite mixture modeling. Results: The estimated DR prevalence in the population represented by Cerner EHR is approximately 25 % and the classification techniques generally achieve an AUC of around 87 %. As a by-product, a predictive inference on the risk of DR based on a patient's personalized medical information is derived. Discussion: Missing labels is a common issue for EHR data quality. Ignoring these missing labels can lead to biased results in the analyses of EHR data. The problem is especially severe in the context of DR. It is thus important to use machine learning or statistical tools to identify the unlabeled patients. The tool in this paper helps both data analysts and clinicians in their practices.