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

This review examines how the selection of data representation in magnetoencephalography (MEG) research aligns with the objectives of machine learning (ML)-based brain state classification. We explore how different research goals—ranging from testing deep neural networks (DNNs) to developing accessible clinical diagnostic tools—shape the choice of data representation and feature extraction methods. Sensor-level signals, when paired with classical ML methods such as linear discriminant analysis (LDA) and support vector machines (SVM), offer computational efficiency and simplified preprocessing, making them advantageous for applications such as brain–computer interfaces (BCIs) and rapid diagnostics. They are also effectively utilised by deep learning (DL) models, such as convolutional neural networks (CNNs), particularly in handling high-dimensional raw data. In contrast, reconstructed source signals provide superior spatial localisation, making them better suited for applications requiring anatomical mapping, such as functional connectivity analysis. These signals are often integrated with DL methods for advanced feature extraction or combined with classical ML techniques to enhance interpretability and robustness in lower dimensional settings. By systematically reviewing these approaches, we highlight how study objectives drive data representation choices and discuss their implications for preprocessing strategies and classification methodologies across different MEG applications.