Investigation on the effectiveness of EEG-BCI techniques to analyse meditation : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Computer Science at Massey University, Manawatū, New Zealand

Abstract

Electroencephalography (EEG), combined with computational approaches such as signal processing and machine learning, has emerged as a powerful tool for assessing meditation states. However, key gaps remain, particularly in analysing EEG data across multiple sessions and evaluating classification performance under both intra-subject and inter-subject (subject-independent) settings. This thesis addresses these challenges through both a systematic computational review and a series of novel experimental studies focused on EEG-based meditation classification. First, a comprehensive review was conducted to systematically organise and evaluate EEG feature extraction and classification techniques used in meditation research. In addition, the review examined EEG preprocessing practices, brain wave frequency usage, and meditation and participant characteristics, providing a valuable computational reference for future researchers in the field. The experimental component of this study used an online EEG dataset collected over multiple sessions for four mental tasks: Loving-Kindness Meditation directed toward the self (LKM-Self), toward others (LKM-Others), Pre-Resting (non-meditation), and Post-Resting. A Brain–Computer Interface (BCI) pipeline using Common Spatial Patterns (CSP) for feature extraction and Linear Discriminant Analysis (LDA) for classification achieved strong performance. Intra-subject, single-session meditation versus non-meditation classification achieved a mean accuracy of 99.5%. Using all four mental tasks, the results showed that Pre-Resting was the most distinguishable condition, while the remaining tasks exhibited more similar neural patterns. The research was then extended to intra-subject, multi-session classification, where pooled session data were used to assess classification consistency over time. These results confirmed the presence of recurring neural patterns across sessions, similar to those observed in single-session analysis, and achieved a mean intra-subject, multi-session meditation versus non-meditation classification accuracy of 83.6%. Building on this, a more advanced task was introduced to classify a new session pair, defined as a meditation session and a non-meditation session selected as a matched pair, after training on other session pairs from the same individual. To support this, following multiple tests to identify the best algorithms, three BCI pipelines were developed using CSP, Short-Time Fourier Transform (STFT), and their fusion (CSP + STFT), also referred to in this thesis as CSP-STFT fusion, with neural networks employed as the consistent classification algorithm. Across different experimental setups, CSP-STFT fusion yielded the highest performance, achieving an overall mean accuracy of 72.9% for intra-subject, multi-session classification. Further analysis showed that increasing the number of training session pairs (from 2 to 4) led to improved classification outcomes. The highest mean accuracy of 75.5% was achieved for LKM-Self versus non-meditation using four training session pairs, confirming the potential to classify new sessions based on prior data. These results demonstrate the feasibility of intra-subject, multi-session EEG classification for previously unseen session pairs. In the final stage, the same three pipelines were evaluated under a subject-independent multi-session classification framework, allowing for direct comparison with intra-subject performance. A total of 9,900 independent tests were conducted to classify new session pairs across individuals. While overall accuracy was lower compared to intra-subject results, substantial variability was observed depending on session combinations. To highlight this, accuracies were divided into top 50% and bottom 50%, and the mean of each group was calculated. For example, in the classification of LKM-Self versus non-meditation using three training session pairs, the CSP + STFT pipeline achieved a mean accuracy of 62.3% overall, with 46.0% for the bottom 50% and 78.3% for the top 50%. This was lower than the intra-subject accuracy of 72.1% under similar conditions, with similar trends observed across other tasks. Based on the mean accuracies obtained, the fusion approach outperformed individual methods in 83.3% of the cases. Additionally, increasing the number of training session pairs led to improved mean accuracies in approximately 75.0% of experiments, emphasizing the importance of training data volume in complex, subject-independent settings. The comparison between intra-subject and subject-independent outcomes confirmed that CSP-STFT fusion enhances classification robustness. Although subject-independent classification remains more challenging due to participant variability, the ability to obtain strong accuracies for certain training combinations is promising. These findings suggest that further research is needed to improve generalizability in real-world applications. These findings provide a strong foundation for developing both personalized and subject-independent EEG-based Brain-Computer Interface (BCI) applications aimed at monitoring and enhancing meditation practice. The demonstrated effectiveness of combining spatial and time-frequency features, along with evidence of recurring neural patterns across sessions, highlights the potential for real-time, adaptive meditation support systems. Additionally, the methodologies and insights presented in this thesis offer valuable directions for future research in scalable BCI pipelines, continuous mental state tracking, and integration with multimodal neuroimaging approaches.