Using a retrospective cohort of over 44,000 ICU patients, we derive a predictive model using machine learning to indicate the onset of circulatory system deterioration. This model integrates data from over 200 continuously-monitored values, including vital signs, lab test results, and treatment parameters. We achieve an AUROC of 0.93 on predicting circulatory system deterioration in the next two hours, and an AUROC of 0.90 for deteriorations up to 6 hours in advance.

Learning Objective 1: After participating in this session, the learner should be better able to:
- Understand the use of early warning systems in the critical care setting, their limitations and current state of the art
- Identify the many challenges of working with large datasets collected from routine clinical practice, and solutions for data preparation
- Understand how machine learning can exploit multivariate time series for real-time deterioration prediction
- Learn which physiological and treatment parameters are most predictive of circulatory system deterioration


Stephanie Hyland (Presenter)
ETH Zurich

Martin Faltys, Bern University Hospital
Matthias Hüser, ETH Zurich
Xinrui Lyu, ETH Zurich
Cristóbal Esteban, ETH Zurich
Gunnar Rätsch, ETH Zurich
Tobias Merz, Bern University Hospital

Presentation Materials: