Interpretability of Sudden Concept Drift in Medical Informatics Domain

Concept drift is usually met in rapidly changing environments, especially in sequential data classification, where different types of concept drift occur on regular basis. This paper presents an approach to dynamic visualization of sequential data characteristics aiming to improve the comprehensibility of concept drifts that result in significant change of classification performance. The proposed approach is applied to sequential multi-label hospital discharge dataset containing diagnosis information for more than two million patients. Our experimental results demonstrate visualization of the anomalies in diagnosis coding through time that can explain the differences in sudden changes of class distribution or classification performance.