Decoding Cognitive Load Changes Induced by Mantra Meditation From Physiological Signals Using Deep Neural Networks

Mantra meditation is widely recognized for enhancing cognition and autonomic regulation; however, its specific impact on cognitive load remains underexplored. Existing research often relies on subjective assessments or conventional electroencephalogram (EEG) spectral analysis, lacking objective, integrated evaluation frameworks. To address these limitations, this study combines physiological signals: EEG-derived cognitive load indices, heart rate variability (HRV) markers, and deep learning-based cognitive state classification to assess the effects of mantra meditation. Cognitive load was quantified using the cognitive load quotient (CLQ) and neural efficiency index (NEI) from EEG signals, while HRV parameters-coherence index (CI) and stress index (SI) were used to evaluate autonomic modulation. Deep learning models, including LSTM, CNN-LSTM, and EEGNet, classified cognitive states (pre-task, task, post-task) across two sessions. In the experimental group, EEGNet achieved accuracies of 98.04% and 91.48%, CNN-LSTM 80.61% and 65.1%, and LSTM 58.29% and 38.74% in sessions 1 and 2, respectively. In contrast, the control group showed relatively stable accuracies: EEGNet (97.46% to 97.98%), CNN-LSTM (81.33% to 83.33%), and LSTM (61.36% to 58.37%). The post-intervention reduction in classification accuracy for the experimental group suggests that decreased cognitive load led to less distinct EEG patterns across phases. Results demonstrated a significant reduction in CLQ and SI, and an increase in NEI and CI, indicating improved cognitive efficiency and autonomic balance. Overall, the findings suggest that mantra meditation fosters neurophysiological efficiency, reflected in both autonomic and neural markers, while reducing the neural separability of cognitive phases-an effect captured through deep learning-based EEG classification.