Electroencephalography classification technique based on statistical denoising and modified k-nearest neighbor algorithm with bipolar sigmoid rectified linear unit’s function

Accurate classification of electroencephalography (EEG) data is much needed for early identification of diseases to treat various disorders. In this paper, we propose EEG classification technique based on statistical denoising & modified k-nearest neighbor (k-NN) algorithm with bipolar sigmoid rectified linear units (ReLU) function. The EEG data is subjected to statistical methods to remove the artifacts and then applied to modified k-NN algorithm to categorize the appropriate features giving preference to neighbors closer to one another considering the weighted votes of the k-nearest neighbors before selecting the class label based on the highest weighted vote. A customized activation function that combines these two functions called as hybrid function that uses various portions of each function in particular ranges is used in our work i.e., use of bipolar sigmoid for negative values and the ReLU function for positive values which helps to limit the signal in a particular range. The proposed algorithm's detection accuracy is tested for the confusion matrix of true positive (TP), false positive (FP), false negative (FN)and true negative (TN) and compared to the detection accuracy of other existing algorithms, demonstrating the algorithm's efficiency with a classification accuracy of almost 85 percent and sensitivity of 91% for standard Kaggle Dataset.