MountPat: investigations on the EEG signals

Abstract

To extract information from the brain, the most cost-effective method is electroencephalography (EEG) signal acquisition. Therefore, many researchers have used EEG signals to capture brain activity. EEG signals are complex; hence, computer-aided models-especially machine learning (ML)-are generally employed to interpret them. The primary objective of this research is to demonstrate the feature-extraction capability of a new, novel method. The proposed feature-extraction approach employs a deterministic feature-engineering transformation, designed to restructure multi-strided signal representations through fixed linear operations. The resulting transformation graph exhibits a mountain-like structure; therefore, we term the model MountPat. To evaluate MountPat's performance, we present an explainable feature engineering (XFE) model with four main phases. In the first phase, we extract informative features using MountPat. In the second phase, we select the most informative features using cumulative weighted iterative neighborhood component analysis (CWNCA). In the third phase, we generate classification results by applying t-algorithm-based k-nearest neighbors (tkNN). In the fourth phase, we extract explainable insights from the EEG signals using the Directed Lobish (DLob) explainable artificial intelligence (XAI) method. To demonstrate the general classification ability of the MountPat-based XFE framework, we use six EEG datasets. Under rigorous subject-independent (LOSO) validation, the model achieves 76.36%-98.88% accuracy, demonstrating strong cross-subject generalization. Sample-wise tenfold CV results exceed 89% on all six datasets. Moreover, by deploying the DLob XAI method, we generate interpretable results for each dataset. These results clearly illustrate that the MountPat-based XFE framework is an effective feature-extraction approach for multichannel signal processing.