OBJECTIVES: Patients enrolled in Veterans Affairs (VA) hospitals are eligible for cost-free medical and pharmacy benefits for conditions developed during military service as determined during compensation and pension exam(s) (Code of Federal Regulations, Chapter 38). As it relates to pharmacy benefits, practitioners point-of-care decisions which determine if first-party and third-party billing events occur. This quality improvement project aims to identify instances of discordance between pharmacy benefits decisions in point-of-care decisions and predictive analysis. METHODS: During a proof-of-concept pilot, prescription records were extracted prospectively to build a repository of probabilities that a particular disability was associated with a medication using logistic regression and automated model selection, stepwise minimization of Akaike information criterion (AIC). The subset of known probabilities were then applied to the VA-wide database for prescriptions to evaluate pharmacy benefit decisions.   RESULTS: During the pilot run approximately 5,742 / 3 million drug-disability pairs were evaluated and the probability of being service connected or non-service connected was applied to 1,026,677 prescriptions with pharmacy benefits determinations during the month of January 2015. Due to the partial database of drug-disability predictions, 317,474 prescriptions were evaluated on 550,239 drug-disability pairs. This method found that 43,571 / 317,474 (13.7%) prescriptions had a pharmacy benefits decision that conflicted with the predicted probability, and 18,185 (5.7%) fell into a high probability that a variance occurred. CONCLUSIONS: Each month a considerable volume of pharmacy benefit decisions may warrant further clinical and/or administrative review. If validated, statistical and machine learning techniques could be a reasonable approach to analyze benefits decisions in VA hospitals.