Application of clustering methods for detecting critical acute coronary syndrome patients

Abstract

We investigate the applicability of two clustering algorithms, DBSCAN and k-means, to detection of critical (died) patients using medical parameter time series. In addition, we perform preliminary cluster analysis of outliers. The most important motivation behind this paper is the potential use of these methods in real clinical setting within an automatic early warning system to potentially decrease the in-hospital mortality rates. Investigation of the outlier clustering is an important step towards the outlier analysis that will help finding out the cause of these outliers: a human error in the records, a technical mistake in the lab, or an actual complication of a patient’s disease. Our results demonstrate that specific clustering algorithms achieve a moderate performance in critical patient detection (F1 scores 0.489 for DBSCAN and 0.495 for k-means). The patient classification boundary is very complex and could not be accurately detected by thesealgorithms, but there are methods capable to determine such boundaries, e.g. the high-degree polynomial kernel SVM. Regarding the outlier clustering, we perform a preliminary analysis showing that it is a viable option for potential outlier classification and analysis.