Deep Learning with Convolutional Neural Network for detecting microsleep states from EEG: A comparison between the oversampling technique and cost-based learning

• Published in: Conference proceedings (IEEE Engineering in Medicine and Biology Society. Conf.) Vol. 2019; pp. 4152 - 4155
• Main Authors: Krishnamoorthy, Venkat, Shoorangiz, Reza, Weddell, Stephen J., Beckert, Lutz, Jones, Richard D.
• Format: Conference Proceeding Journal Article
• Language: English
• Published: United States IEEE 01.07.2019
• Subjects: Attention; Brain modeling; Computer crashes; Convolution; Deep Learning; Electroencephalography; Feature extraction; Humans; Neural Networks, Computer; Sleep Stages; Task analysis; Training
• ISSN: 1557-170X; 1558-4615
• DOI: 10.1109/EMBC.2019.8857588

Any occupation which involves critical decision making in real-time requires attention and concentration. When repetitive and expanded working periods are encountered, it can result in microsleeps. Microsleeps are complete lapses in which a subject involuntarily stops responding to the task that they are currently performing due to temporary interruptions in visual-motor and cognitive coordination. Microsleeps can last up to 15 s while performing a particular task. In this study, the ability of a convolutional neural network (CNN) to detect microsleep states from 16-channel EEG data from 8 subjects, performing a 1D visuomotor was explored. The data were highly imbalanced. When averaged across 8 subjects there were 17 responsive states for every microsleep state. Two approaches were used to handle the CNN training with data imbalance - oversampling the minority class and cost-based learning. The EEG was analysed using a 4-s epoch with a step size of 0.25 s. Leave-one-subject-out cross-validation was used to evaluate the performance. The performance measures used for assessing the detection capability of the CNN were: sensitivity, precision, phi, geometric mean (GM), AUC ROC , and AUC PR . The performance measures obtained using the oversampling and cost-based learning methods were: AUC ROC = 0.90/0.90, AUC PR = 0.41/0.41 and a phi = 0.42/0.40, respectively. Although the performances were similar, the cost-based learning method had a considerably shorter training time than the oversampling method.