for the podium presentation abstract submitted for AMIA 2016
Lujie Chen (1,2) MS, Gilles Clermont(3), MD, Marilyn Hravnak(4), PhD Michael R. Pinsky(4), MD, Artur Dubrawski(2), PhD
(1)Heinz College and (2)School of Computer Science, Carnegie Mellon University, and (3)School of Medicine and (4)School of Nursing, University of Pittsburgh, Pittsburgh, PA
Monitoring cardiorespiratory instability (CRI) in intensive care is challenging, primarily due to the heterogeneity of patterns of CRI risk escalation and the diversity of the involved risk drivers. We present a framework for producing interpretable visualizations of CRI risk estimates obtained from multiclass machine learning models. Our approach can be used as a tool for a “bird’s eye view” of patients’ risk characteristics, and support online monitoring and real time tracking of risk of CRI.
This video shows the risk state evolution with respect to risk esclation pattern (left plot) and risk drivers (right plot) during the patient’s length of stay. The frames are sampled at a rate of once every 15 mins, with a sliding window size of 2 hours. Only data points with risk level over 0.5 are shown in the videos.
This video demos visual tracking for retrospective analysis purpose. The video starts with a small window toward the end of patient’s stay and gradually to include more data from the past as the video progress.