ICU mortality prediction using time series motifs


In this paper, we explore the application of motif discovery (i.e., the discovery of short characteristic patterns in a time series) to the clinical challenge of predicting intensive care unit (ICU) mortality. As part of the Physionet/CinC 2012 challenge, we present an approach that identifies and integrates information in motifs that are statistically over-or under-represented in ICU time series of patients experiencing in-hospital mortality. This is done through a three step process, where ICU time series are first discretized into sequences of symbols (by segmenting and partitioning them into periods of low, medium and high measurements); the resulting sequences of symbols are then searched for short subsequences that are associated with in-hospital mortality; and the information in many such clinically useful subsequences is integrated into models that can assess new patients. When evaluated on data from the Physionet/CinC 2012 challenge, our approach outperformed existing clinical scoring systems such as SAPSII, APACHEII and SOFA, with an event 1 score of 0.46 and an event 2 score of 56.45 on the final test set.

2 Figures and Tables

Cite this paper

@article{McMillan2012ICUMP, title={ICU mortality prediction using time series motifs}, author={Sean McMillan and Chih-Chun Chia and Alexander Van Esbroeck and Ilan Rubinfeld and Zeeshan Syed}, journal={2012 Computing in Cardiology}, year={2012}, pages={265-268} }