0316 Automatic Detection Of Cortical Arousals Using Recurrent Neural Networks

• Published in: Sleep (New York, N.Y.) N.Y.), 2019-04, Vol.42 (Supplement_1), p.A129-A130
• Main Authors: Ragnarsdottir, Hanna, Thrainsson, Heidar Mar, Finnsson, Eysteinn, Gunnlaugsson, Eysteinn, Jonsson, Sigurdur Aegir, Agustsson, Jon Skirnir, Helgadottir, Halla
• Format: Article
• Language: English
• Subjects: Datasets; Neural networks
• ISSN: 0161-8105; 1550-9109
• DOI: 10.1093/sleep/zsz067.315

Introduction Scoring of arousals in EEG signals is a time-consuming and the agreement between experts has been reported with intraclass correlation (ICC) of 0.54 to 0.76. Advances in machine learning and artificial neural networks make it feasible to train computer models to detect arousals in EEG signals. The main challenges, in training the models, are the low agreement between experts, the scoring of arousals is often not accurate since labels are not placed exactly over the arousal, and arousals make up a small portion of the sleep study. Methods We present a method for detecting arousals in EEG signals from polysomnographies (PSG). Features of clinical and statistical origins were derived from the EEG signals and fed into a Bidirectional Recurrent Neural Network, using Long Short-Term Memory units (BRNN-LSTM). The predictions of five neural networks, trained using different features and training sets, were averaged for each sample. The method was developed and validated on two data sets one containing 165 clinical PSGs and the other from the PhysioNet 2018 Challenge dataset consisting of a training set of 994 subjects and a hidden test set of 989 subjects. Results The ICC calculated on a hidden test set of 16 recordings, randomly selected from the clinical PSG studies, was ICC(2,1) 0.88 and the area under precision-recall curve (AUPRC) was 0.81. Five-fold cross-validation on the PhysioNet data set resulted in an ICC(2,1) of 0.59, AUPRC score of 0.45, and on the hidden test set the AUPRC score was 0.45. Conclusion Effectively scoring arousals automatically is important, as manual scoring of arousals is time consuming and difficult. The automatic analysis from the clinical data set showed better results than reported from manuals scoring. The PhysioNet dataset showed lower validation score due to the lower quality of the signals and the manual scoring. Implementing automatic arousal scoring into a commercial software will make new analysis available to the clinic. Support (If Any) This work was supported by the Icelandic Centre for Research under the Icelandic Student Innovation Fund and the Horizon 2020 SME Instrument number 733461.