Classification of Encephalographic Records Using ICA Analysis and Neural Networks

This is the authors' abstract. We don't add key points for this paper.

This article is dedicated to the development and analysis of innovative methods for automated artifact classification in electroencephalographic (EEG) recordings using independent component analysis (ICA) and modern neural network architectures. The aim of the study was to compare the performance of various deep learning models for automatic recognition of the most common types of artifacts, including ocular, cardiac, myographic, and technical noise, which significantly complicate EEG data interpretation in clinical practice. The study considered models based on recurrent neural networks (RNN), LSTM architectures with two-layer structure and dropout = 0.2, convolutional neural networks (CNN with three layers), and transformers with attention mechanism. To ensure the reliability of the results, a consistent data preprocessing strategy was implemented, including signal filtering, class balancing, and optimization of training parameters (Adam optimizer, batch size = 32, train/test split 80/20). The comprehensive approach enabled assessment of the specific characteristics of each artifact class and a detailed error analysis at various stages of signal processing. The best performance was observed in the LSTM architecture (F1-score 0.86), confirming its ability to efficiently model complex temporal dependencies in EEG signals. The CNN model also demonstrated high accuracy (F1-score 0.83), especially for detecting motor patterns, but was less sensitive to prolonged artifacts. The obtained data demonstrate the advantage of deep learning over traditional algorithms such as SVM and linear classification, confirming the possibility of fully automating the EEG preprocessing stage for neurological disease diagnostics. The proposed methods contribute to reducing subjective errors, shortening analysis time, and expanding the prospects for implementing intelligent systems in modern neurophysiological practice, ensuring increased accuracy and versatility of diagnostics.