Managing healthcare costs by peer-group modeling

• Published in: Applied intelligence (Dordrecht, Netherlands) Vol. 43; no. 4; pp. 752 - 759
• Main Authors: Weiss, Sholom M., Kulikowski, Casimir A., Galen, Robert S., Olsen, Peder A., Natarajan, Ramesh
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.12.2015; Springer Nature B.V
• Subjects: Artificial Intelligence; Codes; Computer Science; Cost reduction; Costs; Drug stores; Expected values; Fraud; Fraud prevention; Health care; Health care expenditures; Hypertension; Industrialized nations; Intelligence; Kidney diseases; Machines; Manufacturing; Mathematical models; Mechanical Engineering; Outliers (statistics); Patients; Peers; Pharmacy; Physicians; Processes; Statistical methods; Nearest neighbor; Healthcare; Machine learning; Prediction; Peer profiles
• ISSN: 0924-669X; 1573-7497
• DOI: 10.1007/s10489-015-0685-7

We describe statistical methods for managing healthcare costs using peer-group models and outlier detection. A peer group is a collection of similar entities such as patients, physicians, clinics, hospitals or pharmacies. In an empirical study of drug volumes prescribed by physicians, we examined the billing and prescription records for all patients covered by a major insurer over a 6 month period, encompassing over twenty million individual patient-physician encounters. During this period, 21,243 physicians prescribed a major pain-control medication which is frequently the subject of abuse - oxycodone. Profiles were computed for each physician based on their specialty and the clinical characteristics of their patients. For each physician, the average prescription volume within the corresponding peer group of similar physicians is an estimate of the expected volume of prescriptions for that physician. Strategies were developed to select outliers from the expected values as the ones that are candidates for potential cost reduction. Overall, the prediction of actual outcomes from peer profiles is significantly better than chance, with a reduction of average error of 45.5 %. For the 10 % of physicians that prescribed the most medications, there were extreme and highly significant differences found between their expected and predicted outcomes.