A machine learning pipeline that fuses two feature-extraction methods—Common Spatial Patterns (CSP) and Short Time Fourier Transform (STFT)—classified loving-kindness meditation (LKM) versus non-meditation states from EEG data more accurately than either method alone. Using multiple sessions of EEG from 12 participants, the combined pipeline achieved an overall accuracy of 72.9%, compared to 67.1% for CSP alone and 67.8% for STFT alone. The highest mean accuracy for a single participant was 75.5% for LKM-Self versus non-meditation with five sessions, and the top individual accuracy reached 88.9%. Classification improved as the number of training sessions increased from two to four.
Classifying loving kindness meditation (LKM) from non-meditation using electroencephalography (EEG) data is possible for both single and multiple sessions. For a single session with 32 participants, classification accuracy reached about 99.5% for distinguishing meditation from pre-resting states. For 15 participants across five sessions, accuracy was about 83.6%. Common Spatial Patterns, a feature extraction method typically used in motor imagery brain-computer interfaces, was applied to meditation EEG data for the first time. Comparisons among four mind tasks—pre-resting, post-resting, LKM directed toward self, and LKM directed toward others—showed that pre-resting states were more easily distinguished from the other three, suggesting pre-resting has distinct neural features.
Classifying meditation versus non-meditation from EEG data across different people and multiple sessions is feasible but less accurate than classifying within a single person. Twelve participants each completed five sessions of loving-kindness meditation (directed toward self or others) and non-meditation. Machine learning algorithms trained on session data from a common pool achieved mean accuracies up to 62.3% for distinguishing self-directed meditation from non-meditation, compared to 72.1% in previous within-subject work. Combining two feature extraction methods (common spatial patterns and short-time Fourier transform) outperformed either method alone in 83.3% of test instances, and using more training sessions improved accuracy in 75.0% of instances. Results varied widely depending on which sessions were selected for training and testing.
A commentary argues that a recent proposal to define mindfulness as present-centred bare awareness with an equanimous attitude is too narrow. The author contends that excluding memory, attentional regulation, and temporal continuity reduces a broader psychological capacity to mere features of meditative experience. Drawing on cognitive psychology and contemplative science, the commentary suggests mindfulness is better understood as a temporally extended capacity for flexibly regulating attention and attentional mode, including both intentional orienting and deliberate non-orientation. This formulation preserves the refined definition's strengths while aligning with cognitive mechanisms and the stable functional roles ascribed to mindfulness across contemplative traditions.