0116 ACTIVE ENSEMBLE LEARNING FOR EEG EPOCH CLASSIFICATION

• Published in: Sleep (New York, N.Y.) N.Y.), 2017-04, Vol.40 (suppl_1), p.A43-A43
• Main Authors: Gupta, M, Klerman, EB
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Automation; Classification; Electroencephalography; Sleep; Teaching methods
• ISSN: 0161-8105; 1550-9109
• DOI: 10.1093/sleepj/zsx050.115

Abstract Introduction: Manual classification of epochs, such as for scoring of sleep and detection of artifacts and spindles, is time-consuming and expensive. Automated classification methods often exist but have limited accuracy. An ensemble of such automated classifiers can be used to boost overall accuracy. For acceptable accuracy, however, traditional ensemble learning methods require a large set of manually labeled samples for training. We have developed a novel Active Learning (AL) method that boosts classification accuracy of the ensemble by using only a small selective set of training samples. Methods: A total of 8700 two-second EEG epochs (4.83 hours total) collected during three-minute eyes-open Karolinska Drowsiness testing episodes from nine healthy individuals were manually classified by a RPSGT as artifactual or non-artifactual. Six unique automated artifact detection algorithms, each with individual accuracy ranging from 80 to 91%, were used to form an ensemble input to our AL algorithm to classify the epochs. A set of training samples, totaling 10% of the epochs, was optimally selected using a novel iterative algorithm based upon a generative probabilistic model that results in use of each detector to its maximum capacity. Accuracy (epochs correctly classified/total epochs), false positive rates (FPR) and false negative rates (FNR) from this method were compared with that of eight traditional ensemble supervised/semi-supervised learning methods that used random sets of 10% epochs as training samples and five unsupervised ensemble methods. Results: Our method achieved 97.5% accuracy, 0.5% FPR and 1% FNR, which was considerably better than the best traditional ensemble classifier tested (which was supervised, support vector machine based) that had 94.2% accuracy, 3.2% FPR and 3.5% FNR. Conclusion: Our novel AL algorithm can be can used to enhance accuracy in detecting artifacts in EEG using an ensemble of automated detectors, none of which has high accuracy. Our algorithm is computationally simple and requires very few training samples. This method may be applied also to automated sleep staging or detection of sleep spindles and disordered breathing, where an accuracy boost of even 3% can lead to significant cost reduction. Support (If Any): NIH K24-HL105664 (EBK), P01-AG009975, R01-HL-114088, R01GM105018, R01HL128538, R21HD086392, T32 HL07901.