Motor Imagery Eeg Signal Classification Using Ensembled Cnns For Brain Computer Interface

Brain-Computer Interface(BCI)based on EEG signals is a useful alternative of communication for impaired individuals suffering from movement limitation disease.The brain signals used in BCIs are first collected and analyzed,then translated into commands to carry out desired actions.BCI has been used so far in many application domains ranging from health,security,and sport to virtual reality.However,cerebral task recognition using EEG signals is still a difficult problem that continues to stump researchers.BCI-based on Motor Imagery(MI)requires efficient and fast signal processing to be able to identify the subject’s intention successfully.Feature extraction along with the classification of MI-EEG salient features is judged to be a very complicated task despite the good temporal resolution of the EEG.The EEG signals are variable and have a low signal-to-noise ratio.Traditional Machine Learning proposed in the literature lacks the level of performance needed to be used in real-life applications.Furthermore,the majority of studies using Deep Learning(DL)algorithms execute hyperparameter tuning manually based on expert knowledge only,which is both insufficient and time-consuming.In order to overcome the above-mentioned limitations,this thesis presents an automated end-to-end approach for classifying MI task-related features based on a fusion of the optimum values of three Convolutional Neural Network(CNN1,CNN2,CNN3)models,using the Average Ensemble(AE)strategy.Moreover,we adopted the Bayesian Optimization(BO)algorithm for hyperparameter tuning.The proposed CNN model can handle raw or with minor pre-processing EEG data.Also,it is potent to learn signal dependencies and perform advanced feature extraction,feature selection,and classification within an automated pipeline.We trained and evaluated our method using three different datasets BCI Competition IV-2a,BCI Competition IV-2b,and self-collected HEBUT dataset.Each CNN model was tuned by the mean of the Bayesian Optimization algorithm in order to boost the model performance and reduce time cost.We compared our proposed approach with five traditional Machine Learning by using the following performance metrics;accuracy,precision,recall,F1-score and kappa.The experimental results demonstrate that our approach outperformed the Linear Discriminate Analysis(LDA),Support Vector Machine(SVM),Random Forest(RF),Multi-Layer Perceptron(MLP),and Gaussian Naive Bayes(GNB)algorithms with a wide margin on BCI Competition IV-2a dataset with an accuracy of 92%,and a kappa score of 89.3%,whereas on BCI Competition IV-2b dataset,it attained an accuracy of 89.4%,and a kappa score of 94.5%.Our proposed method on HEBUT dataset attained an accuracy of 94.2%,and a kappa score of92.1%.On the other hand,the GNB algorithm was the weakest performer with an accuracy of41.1%,55% and 53% respectively.Further,the results of our approach outperformed state-ofthe-art techniques on the BCI competition IV-2a multiclass MI database.The outcome of our study showed the benefit of combining the output of Average Ensembled CNN models with Bayesian automated hyperparameter tuning,providing a promising solution for accelerating the EEG based BCI clinical application process.