Aim-based choice of strategy for MEG-based brain state classification

European Physical Journal: Special Topics. 2025. Vol. 234. P. 4159–4177.

Saranskaia I., Boris Gutkin, Zakharov D.

This review explores the interplay between data representation and machine learning (ML) methods in classifying functional, cognitive, and pathological brain states using magnetoencephalography (MEG). Two primary data representations are considered: sensor-level signals and reconstructed source signals. Sensor signals, when combined with classical ML methods such as linear discriminant analysis (LDA) and support vector machines (SVM), offer computational efficiency and simplified preprocessing, making them suitable for applications like
brain-computer interfaces (BCIs) and rapid diagnostics. Sensor signals are also effectively used by advanced deep learning (DL) techniques, such as convolutional neural networks (CNNs), particularly for tasks involving minimally preprocessed, high-dimensional data. On the other hand, source signals, providing superior spatial localisation, are well-suited for tasks requiring anatomical mapping, such as functional connectivity analysis. These signals are often combined with DL approaches for complex feature extraction or paired with classical ML methods to leverage their interpretability and robustness in low-dimensional data scenarios. This review highlights how the strategic combination of data representation and ML methods can optimise preprocessing strategies and enhance classification accuracy across diverse brain state classification tasks.

Keywords: машинное обучение, Magnetoencephalography (MEG), магнитоэнцефалография (МЭГ), Machine learning, brain state classification, классификация состояний мозга

Publication based on the results of: Multidisciplinary study of behavior and decision-making in health population and patients using behavioral, economic, neurocognitive, neuroeconomic, neurocomputational and neural network approaches (2025)