SmartPrognosis: Automatic ensemble classification for quantitative EEG analysis in patients resuscitated from cardiac arrest

Abstract

Understanding a patient’s current status, anticipated clinical course, and likely outcomes can be critical to the practice of medicine. Among patients who are comatose after resuscitation from cardiac arrest, identifying those with potential for awakening and favorable recovery is challenging. Currently, this task is accomplished through the acquisition of one or more diagnostic modalities that aim to assess brain function, with expert interpretation of the test results. This approach is subjective, imprecise, and not scalable. We propose an automatic ensemble classification framework, named SmartPrognosis, to identify comatose post-arrest patients with no recovery potential. SmartPrognosis automatically generates and assembles candidate machine learning pipelines with high sensitivity predicting poor outcomes at a fixed near-zero error rate of misclassifying patients with good outcomes. We demonstrate the effectiveness of SmartPrognosis on real patient data, showing that it over-performs commonly used alternative approaches on all evaluation metrics.