S27. Outcome prediction in postanoxic coma with deep learning. (May 2018)

• Record Type: Journal Article
• Authors: Hofmeijer, Jeannette; Ruijter, Barry J.; Beishuizen, Albertus; Bosch, Frank H.; van Putten, Michel J.; Tjepkema-Cloostermans, Marleen C.
• Is Part Of: Clinical neurophysiology. Volume 129(2018)Supplement 1
• Page Start: e152
• Publication Date: 2018-05
• Subjects: Neurophysiology -- Periodicals; Electroencephalography -- Periodicals; Electromyography -- Periodicals; Neurology -- Periodicals; 612.8
• DOI: 10.1016/j.clinph.2018.04.387
• Languages: English
• ISSNs: 1388-2457

Abstract

Abstract : Introduction: Visual assessment of the electroencephalogram (EEG) by experienced clinical neurophysiologists allows reliable outcome prediction in up to half of all comatose patients after cardiac arrest. We hypothesize that deep neural networks can achieve similar or better performance, while being objective and consistent. Methods: In a prospective cohort study, continuous EEG recordings from comatose patients after cardiac arrest were collected from the intensive care units of two large teaching hospitals. Functional outcome at six months was assessed using the Cerebral Performance Category scale (CPC), dichotomized as good (CPC 1–2) or poor (CPC 3–5). Five-minute artifact-free EEG epochs at 12 and 24 h after cardiac arrest were partitioned into 10 s epochs. We trained a convolutional neural network, using the raw EEG epochs and outcome labels as inputs to predict outcome using data from 80% of the patients. Validation was performed in the remaining 20%. The probability of recovery to good neurological outcome was quantified for each individual patient. Analyses of diagnostic accuracy included receiver operating characteristics and calculation of predictive values at 12 and 24 h. Results: Four hundred and fifty-six patients were included, resulting in 306 and 439 EEGs epochs at 12 and 24 h, respectively. Outcome prediction was most accurate at 12 h, with an area under the ROC curve (AUC) of 0.89 versus 0.81 at 24 h. Poor outcome could be predicted at 12 h with a sensitivity of 62% (95% confidence interval (CI): 45–78%) at false positive rate (FPR) of 0% (CI: 0–14%); good outcome could be predicted at 12 h with a sensitivity of 50% (CI: 29–71%) at a FPR of 5% (CI: 1–18%). Conclusion: Deep learning of raw EEG signals outperforms any previously reported outcome predictor of coma after cardiac arrest, including visual assessment by trained EEG expert. Our approach offers the potential for objective and real-time insight in the prognosis of neurological outcome on a continuous scale, and can provide low-cost expertise at the bedside.