Detection and prevention of unusual events becomes an important issue in highly interconnected and computerized environments, mostly because it is demanding both in terms of human labor and capital investment. Anomaly detection provides a set of techniques that are capable of identifying rare (or in other words anomalous) events in the context of large environments (population of patients, entries in a database, etc.). However, even if the results obtained by anomaly detection would be desirable in many areas (monitoring of patients, detection of an intrusion, etc.), existing approaches are limited in their scope, and possibilities for their improvement are far from being exhausted.
Classical approaches in anomaly detection utilize a set of expert-defined rules or distance measures to detect anomalous events. Apart from the fact that these systems often lack well-defined statistical underpinning, adaptation of the systems to a new environment is impossible without direct interaction with a human. On the other hand, statistical methods rely on past data as opposed to carefully extracted expert knowledge. The benefit of the statistical approach is its adaptability. However, existing anomaly applications focus mostly on anomaly detection in unrestricted environments where anomaly is detected with respect to complete population of past observed examples (patients). This biases the anomalies to events that occur with a low prior probability. To use anomaly detection in clinical settings we are interested in finding anomalies (in outcomes, decisions) with respect to patients who suffer from the same or similar condition. To account for the conditional aspect of anomaly detection we are developing a new conditional anomaly detection framework that first selects a set of examples (patient cases) that are similar to the target patient and only after that we determine the deviation of that patient's outcomes from expected patterns. The new method is designed to complement existing knowledge-based detection methods that are clinically precise, but costly to build.
Funding:
Collaborators:
,
Vice Chair, Department of Biomedical Informatics, Director of Pittsburgh Biomedical Informatics Training Program, University of Pittsburgh School of Medicine
Publications:
M. Hauskrecht, M. Valko, B. Kveton, S. Visweswaram, G. Cooper.
Evidence-based anomaly detection in clinical domains.
To appear in the Proceedings of the Annual American Medical Informatics Association (AMIA) conference , 2007.
The paper introduces and develops a new conditional anomaly detection framework for identifying unusual outcomes or
patient management decisions using a repository of past patient cases. The conditional aspect of the method lets us
focus on the patient specific population and detect anomalies with respect to this population. The paper tests the method
on the Pneumonia PORT dataset and evaluates its ability to detect unusual admission decisions.
M. Valko and M. Hauskrecht . Distance metric learning for conditional anomaly detection. In Proceedings of the
Twenty-First International Florida AI Research Society Conference (FLAIRS 2008), to appear.
J. Mezger, G. F. Cooper, M. Hauskrecht, G. Clermont, S Visweswaran.
Detecting Deviations from Usual Medical Care. Annual American Medical Infortmatics Association (AMIA) Conference
The web page is updated by milos.