T59. EEG artifacts removal using machine learning algorithms and independent component analysis

Abstract

Introduction Elimination of electroencephalogram (EEG) artifacts is an important process in EEG analysis and research because artifacts can disturb EEG signals, resulting in serious misinterpretation. Independent component analysis (ICA) is a statistical method to separate independent sources from multivariate data. It has been used to identify and remove the artifacts from raw EEG signals. Topographic maps and power spectrum resulted from ICA show different characteristics for EEG artifact types such as muscle and ocular artifacts. One problem with ICA is that it needs visual inspection by EEG experts. In this study, a new method was developed for automatically classifying and eliminating EEG artifacts using image recognition of topographic maps using the machine learning algorithms. Methods The EEGs recorded from 841 healthy subjects were used in this study. The pipeline for EEG artifacts removal consists of pre-processing including filtering, common average referencing and ICA. Data for training and testing the machine learning algorithms were obtained from manual inspection of the topographic maps and power spectrum by EEG experts. Since the volume of data was important for the machine learning algorithms, generation of additional data was implemented as follows. The pre-processed signals were divided into several segments to extend data for deep learning algorithms. ICA was then applied to the segments and topographic maps and power spectrum for independent components were generated. Artificial neural network (ANN), one of the machine learning algorithms was then used to classify neural signal and 3 different artifacts including EMG, horizontal eye movements (eye blinks), and vertical eye movements. Results The ANN applied to the topographic maps obtained from ICA studies presented its effectiveness to automatically classify and remove the EEG artifacts. An accuracy rate of 91.01 ± 5.12% was obtained and eye blinks were most effectively classified with a recognition rate of 98.29 ± 4.76%. The error rates of missing neural signals and artifacts were 9.19 ± 1.23% and 7.65 ± 2.63% respectively. Finally, the execution time of the system was acquired for real-time applicability. Conclusion A method of the machine learning algorithms applied to ICA has been presented to identify and remove EEG artifacts. The results demonstrate that the proposed method can effectively, automatically remove muscle and ocular artifacts in EEG signals. This study supports that the deep learning is capable of strong potential for removing EEG artifacts resulting high quality of EEG signal analysis without manual inspection by EEG experts. Funding: This work was supported by Ministry of Science and ICT Grant 20170010980011001 .