Abstract:
|
Analytics and predictive inference using electronic health records are trending in recent research for health informatics, decision support, health economics and many. The EHR data are typically very large and thus are considered richly informative for prediction. However, the EHR data are also notoriously noisy due to erroneous inputs, imprecise measurements, or other reasons. Our recent research shows that EHR fitted predictive models/algorithms, though having high AUC in self-validation, face difficulties when predicting real validation data. This calls for adjustments or calibration of EHR-based results by small but reliable clinical information. In this paper, we consider a Bayesian meta-analysis using multiple unbalanced data sources to create a more reliable predictive inference. We apply the proposed approach on a project for early detection of diabetic retinopathy in rural diabetic population.
|