Multivariate Conditional Anomaly Detection for Clinical Anomaly Detection

This paper overviews the background, goals, past achievements and future directions of our research that aims to build a multivariate conditional anomaly detection framework for the clinical application. Background and Goals We humans are prone to error. Despite startling advances in medicine, the occurrence of medical errors remains a persistent and critical problem. Although various computeraided monitoring devices support medical practices to prevent errors, because those tools are primarily knowledgebased built by clinical experts, they are expensive and their clinical coverage is incomplete. We develop a new detection framework that identifies statistically anomalous patient care patterns based on past clinical information stored in an electronic health record (EHR) systems. Our hypothesis is that the detection of anomalies in patient care patterns corresponds to identifying cases that need medical attention for reconsideration. Typical anomaly detection methods, however, simply attempt to identify unusual data instances that do not conform with the majority of examples in the dataset, and are not suitable in the clinical context. This is because clinical decisions on patient care are strongly based on the condition of the patient (Hauskrecht et al. 2013). In addition, patient care generally consists of multiple clinical actions which often show correlations between the individual actions (e.g., a set of medications that are usually ordered together). However, such correlations have not been vigorously exploited in the context of anomaly detection. Our framework aims to improve the anomaly detection performance by identifying multivariate conditional anomalies where we are interested in the patterns exhibit dependencies among individual clinical actions conditioned on the patient condition.