A Taxonomic Framework Based on Machine Learning for Unsupervised Anomaly Detection in Healthcare Data

Objectives: This study presents a generic taxonomic framework as a roadmap for users who wish to implement systems for detecting fraud and waste in health data; this roadmap outlines which unsupervised machine learning techniques can be used to detect anomalies in health data, and whether these anomalies are point, contextual, or collective. Methods: The research employed the Design Science Research (DSR) methodology in five stages: awareness, ideation, development, evaluation, and conclusion. The proposed taxonomic framework was designed through the implementation of the artifact, using the "Sample 1 to Inpatient Claims CMS 2008-2010 Data Entrepreneurs' Synthetic Public Use File (DE-SynPUF)" database, which contains 66,773 records and 81 features. The dataset was applied to the K-Means, Isolation Forest (IF), and Local Outlier Factor (LOF) algorithms for point anomaly detection. Results: The number of anomalies detected by at least one of the algorithms was 13,966, representing (20.9%) of the dataset; K-Means alone detected 1608 (2.4%), IF alone detected 4522 (6.8%), LOF alone detected 2301 (3.4%), K-Means and IF together detected 4497 (6.7%), K-Means and LOF together detected 41 (0.1%), IF + LOF combined detected 465 (0.7%), and all three algorithms in consensus detected 532 (0.8%). Conclusions: Thus, the taxonomic framework proposed in this study can be viewed as a roadmap for researchers and developers seeking to understand which unsupervised technique to apply in detecting anomalies in health data, considering the type of anomaly.