Deep Learning for outcome prediction of postanoxic coma

Abstract

Electroencephalography (EEG) is increasingly used to assist in outcome prediction for patients with a postanoxic coma after cardiac arrest. Current literature shows that neurological outcome is invariably poor if the EEG remains iso-electric or low-voltage at 24 h after cardiac arrest or if it shows burst-suppression with identical bursts; such patterns are observed in approximately 30-50% of patients. Return of continuous EEG rhythms within 12 h after cardiac arrest predicts good neurological outcome with sensitivities in the range of 30 to 50% at specificities near 100%. In previous work, we reported on the Cerebral Recovery Index to assist in the visual assessment of the EEG. In this paper, we explore a deep learning approach, using a convolutional neural network for outcome prediction in patients with a postanoxic encephalopathy. Using EEGs from 287 patients at 12 h after cardiac arrest and 399 patients at 24 h after cardiac arrest, we trained and validated a convolutional neural network with raw EEG data (18 channels, longitudinal bipolar montage). As the outcome measure, we used the Cerebral Performance Category scale (CPC), dichotomized between good (CPC score 1-2) and poor outcome (CPC score 3-5). Using 5 minute artifact-free epochs from the continuous EEG recordings partitioned into 10 s snippets, we trained the convolutional neural network using 80% of the patients. Validation was performed with EEGs from the remaining 20% of patients. Outcome prediction was most accurate at 12 h after cardiac arrest, with a sensitivity of 58% at a specificity of 100% for the prediction of poor outcome. Good neurological outcome could be predicted at 12 h after cardiac arrest with a sensitivity of 58% at a specificity of 97%. In conclusion, we present a classifier for the prediction of neurological outcome after cardiac arrest, based on a convolutional neural network, providing reliable and objective prognostic information.