O-18 Outcome prediction in postanoxic coma with deep learning. Issue 7 (July 2019)

• Record Type: Journal Article
• Authors: Tjepkema-Cloostermans, Marleen C.; Silva Lourenço, Catarina da; Ruijter, Barry J.; Tromp, Selma C.; Drost, Gea; Kornips, Francois H.M.; Beishuizen, Albertus; Bosch, Frank H.; Hofmeijer, Jeannette; Putten, Michel
• Is Part Of: Clinical neurophysiology. Volume 130:Issue 7(2019:Jul.)
• Page Start: e27
• Publication Date: 2019-07
• Subjects: Neurophysiology -- Periodicals; Electroencephalography -- Periodicals; Electromyography -- Periodicals; Neurology -- Periodicals; 612.8
• DOI: 10.1016/j.clinph.2019.04.334
• Languages: English
• ISSNs: 1388-2457

Abstract

Abstract : Background: Visual assessment of the electroencephalogram (EEG) by experienced neurophysiologists allows reliable outcome prediction of approximately half of all comatose patients after cardiac arrest. Deep neural networks hold promise to achieve similar or even better performance, while being more objective and consistent. Material and methods: We performed a prospective cohort study in 5 Dutch hospitals, and included 895 consecutive comatose patients after cardiac arrest. Continuous EEG was recorded during the first 3 days. Functional outcome at six months was dichotomized as good (Cerebral Performance Category, CPC 1–2) or poor (CPC 3–5). We trained a convolutional neural network to predict outcome using EEG data at 12 and 24 h after cardiac arrest. Data from two hospitals were used for training and internal validation ( n = 661). Eighty percent of these data was used for training and cross-validation, the remaining 20% for independent internal validation. Data from the other three hospitals were used for external validation ( n = 234). Results: Prediction of poor outcome was most accurate at 12 h, with a sensitivity in the external validation set of 58% (95% confidence interval (CI): 51–65%) at false positive rate (FPR) of 0% (CI: 0–7%). Good outcome could be predicted at 12 h with a sensitivity of 48% (CI: 45–51%) at a FPR of 5% (CI: 0–15%) in the external validation set. Conclusions: Deep learning of EEG signals outperforms any previously reported outcome predictor of coma after cardiac arrest, including visual EEG assessment. Our approach offers the potential for objective and real-time, bedside insight in the neurological prognosis.