Authors:
James Kemp
1
;
Chris Barker
2
;
Norm Good
3
and
Michael Bain
1
Affiliations:
1
School of Computer Science and Engineering, University of New South Wales, Building K17 UNSW Sydney, Kensington NSW, Australia
;
2
Provider Benefits Integrity Division, Australian Government Department of Health, L10 260 Elizabeth Street, Surry Hills NSW, Australia
;
3
Australian eHealth Research Centre, Commonwealth Scientific and Industrial Research Organisation, Level 7 STARS Building - Surgical Treatment and Rehabilitation Service 296 Herston Road, Herston QLD, Australia
Keyword(s):
Unsupervised Machine Learning, Data Mining, Orthopedic Procedures, National Health Insurance, Fraud.
Abstract:
Medical fraud and waste is a costly problem for health insurers. Growing volumes and complexity of data add challenges for detection, which data mining and machine learning may solve. We introduce a framework for incorporating domain knowledge (through the use of the claim ontology), learning claim contexts and provider roles (through topic modelling), and estimating repeated, costly behaviours (by comparison of provider costs to expected costs in each discovered context). When applied to orthopaedic surgery claims, our models highlighted both known and novel patterns of anomalous behaviour. Costly behaviours were ranked highly, which is useful for effective allocation of resources when recovering potentially fraudulent or wasteful claims. Further work on incorporating context discovery and domain knowledge into fraud detection algorithms on medical insurance claim data could improve results in this field.