NeuroState: A Comparative Framework for Electroencephalography-Based Meditative State Classification

Accurate identification of meditative states using electroencephalography (EEG) is crucial for objective neural assessment and brain-computer interface applications.However, reliable classification remains challenging because of signal nonstationarity, artifacts, inter-individual variability, and differences across recording sessions.To address these challenges, NeuroState, a dual-path classification framework integrating time-frequency analysis, conventional machine learning, and attention-enhanced deep learning, was developed for automated classification of meditation-related cognitive states.Initially, EEG signals were transformed into time-frequency representations using the Morlet continuous wavelet transform (CWT), allowing temporal variations in spectral activity to be characterized.Two complementary classification pathways were subsequently implemented.In the conventional machine-learning pathway, wavelet-derived scalograms were vectorized, and dimensionality was reduced using principal component analysis (PCA) before classification with support vector machine and random forest classifiers.In the deep-learning pathway, the two-dimensional scalograms were directly processed using a wide squeeze-and-excitation residual network, in which residual feature extraction and channel-wise attention were integrated to facilitate the adaptive learning of discriminative time-frequency representations.The framework was evaluated using the publicly available OpenNeuro ds001787 dataset, comprising EEG recordings from 24 participants, including 12 experienced meditators and 12 individuals with limited meditation experience.The wide squeeze-and-excitation residual network achieved a test accuracy of 98.29%, outperforming the support vector machine and random forest classifiers, which achieved accuracies of 88.59% and 81.65%, respectively.These findings demonstrate the superior classification performance of attention-enhanced deep learning compared with the evaluated conventional machine-learning approaches.These findings suggest that attention-enhanced deep learning applied to structured EEG time-frequency representations may provide more accurate discrimination of meditation-related cognitive states than conventional machine-learning approaches based on PCA-reduced features.NeuroState thus provides a promising computational framework for EEG-based meditative state classification, with potential applications in meditation monitoring, neurofeedback, and brain-computer interface systems.