Classification of meditative phases through EEG analysis

Technical Summary: Classification of Meditative Phases through EEG Analysis

Problem Statement
Meditation is widely recognized for psychological benefits, yet providing objective, subject-independent markers for distinct meditative phases remains a significant challenge. While EEG has been used to detect mind-wandering or distinguish broad states, discriminating between closely related contemplative phases—specifically Concentrative (C), Loving-Kindness (LK), and Emptiness (E)—is difficult. Existing approaches often struggle with cross-subject generalization, where models trained on one group fail to classify data from new subjects. Furthermore, high-performing deep learning models often lack interpretability, making it difficult to derive actionable feedback rules for practitioners. This study addresses the need for a subject-independent classification framework that balances predictive performance with full interpretability.

Methodology
The study utilized a dataset of 35 subjects (30 successfully processed after artifact removal) from Sera Jey Monastery, India. Each participant underwent a protocol involving three 15-minute meditative blocks (C, LK, E) in a fixed order, recorded via a 19-channel EEG system (10-20 standard, 512 Hz downsampled to 100 Hz).

1. Data Representation and Preprocessing:
Unlike traditional approaches that treat short time windows as independent samples, the authors proposed a temporal-spatial representation:

• Feature Extraction: From 15-minute recordings, features were extracted at a 1-minute resolution. The feature set included 601 features comprising:
  • 16 global features (spectral entropy, band power, coherence, phase lag index).
  • 45 local Region-of-Interest (ROI) features (power per region/band).
  • 540 inter-regional connectivity features (real coherence, imaginary coherence, and phase lag index across 36 region pairs and 5 frequency bands).
• Temporal Derivative Encoding: To capture slow dynamics, the 15-point time series for each feature was transformed into a first-order derivative. These derivatives were discretized into three categorical trends: "decrease" (< -1 SD), "stability" (-1 to 1 SD), and "increase" (> 1 SD).
• Encoding: The categorical trends were converted into fixed-length binary vectors using an inverse one-hot (zero-hot) scheme, resulting in a 42-dimensional binary vector per original feature (14 time steps × 3 categories). The final input dimension was 25,242 features (601 features × 42).

2. Classification Protocols:

• Validation: The primary evaluation used a Leave-One-Subject-Out (LOSO) protocol to ensure subject-independent generalization. Results were averaged over 10 repetitions with different random seeds.
• Models:
  • Decision Trees (DT): Trained with a maximum depth of 4 to prevent overfitting and ensure human-readable rules. Splitting was based on the Gini Index.
  • Random Forest (RF): An ensemble of 2000 trees with the same depth constraint, used to benchmark performance against the single DT.
  • Baselines: The study compared the proposed method against standard baselines including custom temporal/spectral features with Neural Networks (NN) and Decision Trees, established deep learning architectures (EEGNet, Deep4Net, ShallowFBCSPNet), and a foundation model (LaBraM).

Key Contributions

1. Temporal-Spatial EEG Representation: A novel pipeline that encodes minute-wise feature trajectories and their derivative trends into discrete categorical patterns, preserving temporal structure rather than treating windows as independent.
2. Explainable-by-Design Pipeline: The implementation of a Gini-index decision tree framework specifically designed to yield interpretable decision rules under strict subject-disjoint validation.
3. Identification of Discriminative Signals: Empirical evidence that low-frequency inter-regional connectivity, specifically delta-band coherence between mid-anterior (MA) and left-central (LC) regions during the early phase of meditation (minutes 4–5), serves as the most stable discriminative signal across subjects.

Results

• Baseline Performance: Standard baselines (Deep Learning, Custom Features, Foundation Models) achieved high accuracy (58–88%) under stratified 5-fold cross-validation (5FCV) but degraded to near-chance levels (29–35%) under subject-grouped validation (G5FCV), confirming the difficulty of cross-subject generalization.
• Decision Tree Performance: The proposed DT approach achieved a mean accuracy of 44.67% (±1.6) and a macro-F1 of 43.8% under LOSO. This is significantly above the 33.33% chance level for three classes.
• Random Forest Performance: The RF model achieved a higher mean accuracy of 52.44% (±1.4) and macro-F1 of 51.57% using only coherence-based features.
• Feature Importance: Analysis of the decision trees and RF impurity scores consistently highlighted local coherence features as the most informative. The most recurrent root node feature was the coherence between MA and LC regions in the delta band at time index t3 (transition between minutes 4 and 5).
• Inter-Subject Variability: While some subjects were classified with near-ceiling accuracy (up to 100%), others remained close to chance, indicating that coherence dynamics are highly informative for specific individuals but not universally uniform.

Significance and Claims
The paper claims that temporally structured connectivity patterns encode reproducible, cross-subject signatures of closely related contemplative phases. The authors argue that while deep learning models offer higher predictive power, they often lack the mechanistic insight required for feedback tools. In contrast, the proposed simple, explainable models provide a principled foundation for future subject-independent feedback systems.

The study emphasizes that early-phase delta coherence is not merely a presence marker but a discriminative feature capable of separating distinct meditative states across different individuals. The authors conclude that their approach successfully demonstrates that cross-subject meditation decoding can be both above chance and interpretable when the representation prioritizes temporal trends of spatial connectivity over short, non-overlapping windows.

Limitations
The authors acknowledge limitations, including the relatively small sample size (35 subjects) and the fixed ordering of meditation phases, which may confound phase effects with time-on-task or fatigue. They note that the fixed sequence is a methodological necessity for preserving the ecological integrity of the Tibetan contemplative practice studied. Future work is proposed to expand to larger cohorts and explore hybrid modeling strategies.