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 

 

 

 

Nalinda Devaka Liyanagedera 

2025 

 



i 
 

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 



ii 
 

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-



iii 
 

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. 

 

  



iv 
 

Acknowledgements 
 

First and foremost, I would like to express my deepest gratitude to my supervisory panel for 

their invaluable guidance, support, and encouragement throughout my PhD journey. I am 

especially thankful to my primary supervisor, Professor Hans W. Guesgen, Chair in Computer 

Science, for the opportunity to study under his supervision. His deep expertise in artificial 

intelligence, along with his exceptional qualities of professionalism, empathy, and kindness, 

and his prompt and encouraging responses, have provided me with both academic inspiration 

and personal encouragement. I am particularly grateful for the freedom he offered me to explore 

interdisciplinary connections beyond his own field, enabling me to integrate computer science 

with meditation and psychological research. 

I am also thankful to my co-supervisors for their expertise and support. Dr Corinne A. Bareham, 

an expert in electroencephalography and cognitive neuroscience, provided essential guidance 

in those areas, while Associate Professor Heather Kempton, whose expertise lies in meditation 

and psychological science, offered invaluable insights that enriched my work. Dr Sunil Lal gave 

me foundational support during the early stages of my doctoral studies. The interdisciplinary 

nature of my research was greatly strengthened by the combined expertise in both computer 

science and psychology on my supervisory team. 

My sincere appreciation extends to the academic, technical, and administrative staff of Massey 

University. I am especially thankful to Professor Chris Scogings, Associate Professor Catherine 

Whitby, Associate Professor Anuradha Mathrani, and Janene de Ridder, whose support helped 

me navigate the unforeseen challenges brought about by the COVID-19 pandemic. I am also 

grateful for their assistance with financial and logistical support, particularly relating to journal 



v 
 

publication costs. I would like to thank all my PhD colleagues and friends at Massey, including 

Ali Abdul Hussain and Amardeep Singh, who made this journey more enjoyable and enriching. 

The EEG meditation dataset used in this research was obtained from an open-access online 

source, originally collected by Ven. GoonFui Wong, Junling Gao, and Rui Sun from the Faculty 

of Education at the University of Hong Kong (HKU). I sincerely appreciate their contribution 

and the availability of this dataset to the wider research community. 

I gratefully acknowledge the financial support I received through a scholarship from the 

Accelerating Higher Education Expansion and Development (AHEAD) project, an initiative of 

the Sri Lankan government funded by the World Bank. 

I also extend my heartfelt thanks to the Vice-Chancellor, Dean, Head of Department, and all 

academic, administrative, and non-academic staff of Wayamba University of Sri Lanka, where 

I am employed as a Senior Lecturer and granted study leave to pursue this PhD. I am thankful 

to my colleagues at Wayamba University for their encouragement and ongoing support 

throughout my studies. 

On a personal note, I am forever grateful to my wife, Ruwanthi Premathilaka, whose 

unwavering support, patience, understanding, and encouragement sustained me through the 

many challenges of this journey, especially as she continues her own PhD in Food Science and 

Technology. I also extend my heartfelt thanks to my mother and father, my brother and his 

family, and my in-laws and their family for their constant love, blessings, sacrifices, and 

encouragement. All of these have been essential to the completion of this work. At the same 

time, my brother Chamika and his wife Parami both completed their PhDs, and their 

achievements provided valuable motivation during my own doctoral journey. 

I would also like to thank all my teachers from school, Trinity College Kandy, and my 

undergraduate studies at the University of Peradeniya, whose early guidance laid the foundation 



vi 
 

for my academic pursuits. In addition, I sincerely thank the supervisory team of my wife for 

their support and guidance throughout her doctoral journey, which indirectly contributed to the 

strength and resilience of our shared academic path. 

This thesis would not have been possible without the collective support of all those mentioned 

above. To each of you, I offer my sincere and enduring gratitude. Finally, I would like to 

conclude with the following words: “Tathagata-pavedito dhammo vinayo vivato virocati, no 

paticchanno.” (The teaching and discipline proclaimed by the Buddha shine when open, not 

when hidden.) 

 

 

  



vii 
 

Table of Contents 
Abstract ........................................................................................................................................ i 

Acknowledgements ................................................................................................................... iv 

List of Figures ............................................................................................................................. x 

List of Tables ........................................................................................................................... xiii 

Chapter 1 Introduction ................................................................................................................ 1 

1.1 Background and Motivation ....................................................................................... 1 

1.2 Problem Statement ...................................................................................................... 6 

1.3 Research Objectives ................................................................................................... 7 

1.4 Thesis Organization .................................................................................................... 9 

Chapter 2 A Computational Review of Feature Extraction and Classification Techniques for 

EEG-Based Meditation Assessment ......................................................................................... 12 

2.1 Introduction .............................................................................................................. 13 

2.1.1. Background on Meditation and EEG Neuroimaging ....................................... 13 

2.1.2. Overview Objectives of the Review ................................................................. 15 

2.2 Methodology ............................................................................................................. 16 

2.2.1. Overview of the Systematic Literature Review Approach ............................... 16 

2.2.2. Research Questions (RQ) ................................................................................. 16 

2.2.3. Search Strategy ................................................................................................. 17 

2.2.4. Inclusion and Exclusion Criteria ...................................................................... 18 

2.2.5. Screening and Selection Process ...................................................................... 18 

2.2.6. Data Extraction and Categorization .................................................................. 19 

2.2.7. Data Synthesis and Analysis ............................................................................. 21 

2.2.8. Limitations of the Methodology ....................................................................... 22 

2.3 Meditation Study Characteristics and Preprocessing ............................................... 23 

2.3.1. Meditation Study Characteristics...................................................................... 23 

2.3.2. Preprocessing .................................................................................................... 25 

2.4 Feature Extraction Techniques for Meditation EEG Assessment ............................. 27 

2.4.1. Introduction of Feature Extraction Techniques ................................................ 27 

2.4.2. Frequency Domain Analysis ............................................................................. 28 

2.4.3. Time-Frequency and Spectral Entropy Methods .............................................. 31 

2.4.4. Amplitude Features .......................................................................................... 33 

2.4.5. Dimensionality Reduction and Decomposition ................................................ 34 

2.4.6. Phase and Synchrony Analysis ......................................................................... 37 

2.4.7. Entropy-Based Measures .................................................................................. 40 



viii 
 

2.4.8. Connectivity and Coherence Measures ............................................................ 42 

2.4.9. Spatial, Microstate, Network and Topological Features ................................... 44 

2.4.10. Statistical Features ............................................................................................ 47 

2.4.11. Advanced Modelling, Machine Learning, and Complex Features ................... 50 

2.4.12. Summary of Feature Extraction Techniques..................................................... 53 

2.5 Classification Techniques for Meditation EEG Assessment .................................... 54 

2.5.1. Introduction of Classification Techniques ........................................................ 54 

2.5.2. One Classification Method ............................................................................... 55 

2.5.3. Two Classification Methods ............................................................................. 61 

2.5.4. Three or More Classification Methods ............................................................. 63 

2.5.5. Summary of Classification Techniques ............................................................ 65 

2.6 Conclusion ................................................................................................................ 67 

2.6.1. Recap of the Review’s Major Insights .............................................................. 67 

2.6.2. Mapping Findings to Current and Future Meditation EEG Studies ................. 68 

Chapter 3 Common spatial pattern for classification of loving kindness meditation EEG for 

single and multiple sessions ................................................................................................... 116 

3.1 Introduction ............................................................................................................ 117 

3.1.1. EEG BCI background ..................................................................................... 118 

3.1.2. EEG signal processing with machine learning techniques ............................. 119 

3.1.3. Study of meditation using EEG ...................................................................... 122 

3.1.4. Summarizing the importance of this study ..................................................... 124 

3.2 Methods .................................................................................................................. 125 

3.2.1. Dataset description ......................................................................................... 125 

3.2.2. Methods of experiment ................................................................................... 127 

3.3 Results .................................................................................................................... 133 

3.4 Discussion ............................................................................................................... 138 

3.4.1. Understanding the EEG meditation dataset based on the results obtained..... 139 

3.4.2. Limitations of the study due to dataset characteristics ................................... 142 

3.5 Conclusion .............................................................................................................. 143 

Chapter 4 Novel machine learning-driven comparative analysis of CSP, STFT, and CSP-STFT 

fusion for EEG data classification across multiple meditation and non-meditation sessions in 

BCI pipeline ............................................................................................................................ 146 

4.1 Introduction ............................................................................................................ 147 

4.1.1. Background of the study ................................................................................. 147 

4.1.2. Related work ................................................................................................... 149 

4.1.3. Meditation in the study ................................................................................... 153 



ix 
 

4.1.4. Importance of this study ................................................................................. 154 

4.2 Methods .................................................................................................................. 155 

4.2.1. Dataset description and overview ................................................................... 155 

4.2.2. Data preprocessing and cleaning .................................................................... 157 

4.2.3. BCI pipeline modelling .................................................................................. 158 

4.2.4. Varied number of sessions .............................................................................. 160 

4.2.5. Feature extraction and classification .............................................................. 162 

4.3 Results .................................................................................................................... 166 

4.4 Discussion ............................................................................................................... 180 

4.4.1. Analysis based on the type of algorithm used ................................................ 182 

4.4.2. Analysis based on the number of training sessions used ................................ 185 

4.4.3. Limitations of the study and future work ....................................................... 198 

4.5 Conclusion .............................................................................................................. 199 

Chapter 5 Machine learning-based comparative analysis of subject-independent EEG data 

classification across multiple meditation and non-meditation sessions ................................. 201 

5.1 Introduction ............................................................................................................ 202 

5.2 Methods .................................................................................................................. 207 

5.3 Results .................................................................................................................... 213 

5.4 Discussion ............................................................................................................... 222 

5.5 Limitations .............................................................................................................. 234 

5.6 Conclusion .............................................................................................................. 235 

Chapter 6 Conclusion and Future Work ................................................................................. 237 

6.1 Summary of the Research Work ............................................................................. 237 

6.2 Key Findings and Contributions ............................................................................. 240 

6.3 Limitations of the Study ......................................................................................... 245 

6.4 Implications and Applications ................................................................................ 248 

6.5 Future Work ............................................................................................................ 249 

References .............................................................................................................................. 252 

 

  



x 
 

List of Figures 
 

Figure 2.1 PRISMA flow diagram for selecting articles for the systematic review ................. 19 

Figure 3.1 Comparison of average of first three accuracy columns (Average accuracy of pairs 

with Pre-Resting) with average of last three accuracy columns (Average accuracy of 

pairs without Pre-Resting) for the 32 participants in Table 3.2 for single session 

classification ............................................................................................................... 137 

Figure 3.2 Comparison of average of first three accuracy columns (Average accuracy of pairs 

with Pre-Resting) with average of last three accuracy columns (Average accuracy of 

pairs without Pre-Resting) for the 15 participants in Table 3.3 for multiple session 

classification ............................................................................................................... 137 

Figure 4.1 Summary of the methodology for EEG data classification across multiple 

meditation and non-meditation sessions ..................................................................... 166 

Figure 4.2 Comparison of average classification accuracies for LKM-Self/ non-meditation 

EEG data when using the three algorithms CSP, STFT and (CSP + STFT) for the case 

of 3 sessions of data .................................................................................................... 169 

Figure 4.3 Relative comparison of normalised average classification accuracies for LKM-Self/ 

non-meditation EEG data when using the three algorithms CSP, STFT and (CSP + 

STFT) for the case of 3 sessions of data ..................................................................... 169 

Figure 4.4 Comparison of average classification accuracies for LKM-Others/ non-meditation 

EEG data when using the three algorithms CSP, STFT and (CSP + STFT) for the case 

of 3 sessions of data .................................................................................................... 171 

Figure 4.5 Relative comparison of normalised average classification accuracies for LKM-

Others/ non-meditation EEG data when using the three algorithms CSP, STFT and 

(CSP + STFT) for the case of 3 sessions of data ........................................................ 171 

Figure 4.6 Comparison of average classification accuracies for LKM-Self/ non-meditation 

EEG data when using the three algorithms CSP, STFT and (CSP + STFT) for the case 

of 4 sessions of data .................................................................................................... 173 

Figure 4.7 Relative comparison of normalised average classification accuracies for LKM-Self/ 

non-meditation EEG data when using the three algorithms CSP, STFT and (CSP + 

STFT) for the case of 4 sessions of data ..................................................................... 173 

Figure 4.8 Comparison of average classification accuracies for LKM-Others/ non-meditation 

EEG data when using the three algorithms CSP, STFT and (CSP + STFT) for the case 

of 4 sessions of data .................................................................................................... 175 

Figure 4.9 Relative comparison of normalised average classification accuracies for LKM-

Others/ non-meditation EEG data when using the three algorithms CSP, STFT and 

(CSP + STFT) for the case of 4 sessions of data ........................................................ 175 

Figure 4.10 Comparison of average classification accuracies for LKM-Self/ non-meditation 

EEG data when using the three algorithms CSP, STFT and (CSP + STFT) for the case 

of 5 sessions of data .................................................................................................... 177 



xi 
 

Figure 4.11 Relative comparison of normalised average classification accuracies for LKM-

Self/ non-meditation EEG data when using the three algorithms CSP, STFT and (CSP 

+ STFT) for the case of 5 sessions of data ................................................................. 177 

Figure 4.12 Comparison of average classification accuracies for LKM-Others/ non-meditation 

EEG data when using the three algorithms CSP, STFT and (CSP + STFT) for the case 

of 5 sessions of data .................................................................................................... 179 

Figure 4.13 Relative comparison of normalised average classification accuracies for LKM-

Others/ non-meditation EEG data when using the three algorithms CSP, STFT and 

(CSP + STFT) for the case of 5 sessions of data ........................................................ 179 

Figure 4.14 Comparison of average classification accuracies for LKM-Self/ non-meditation 

EEG data when using the algorithm CSP for the three cases of 3 sessions, 4 sessions 

and 5 sessions of data ................................................................................................. 187 

Figure 4.15 Relative comparison of normalised average classification accuracies for LKM-

Self/ non-meditation EEG data when using the algorithm CSP for the three cases of 3 

sessions, 4 sessions and 5 sessions of data ................................................................. 187 

Figure 4.16 Comparison of average classification accuracies for LKM-Self/ non-meditation 

EEG data when using the algorithm STFT for the three cases of 3 sessions, 4 sessions 

and 5 sessions of data ................................................................................................. 189 

Figure 4.17 Relative comparison of normalised average classification accuracies for LKM-

Self/ non-meditation EEG data when using the algorithm STFT for the three cases of 3 

sessions, 4 sessions and 5 sessions of data ................................................................. 189 

Figure 4.18 Comparison of average classification accuracies for LKM-Self/ non-meditation 

EEG data when using the algorithm (CSP + STFT) for the three cases of 3 sessions, 4 

sessions and 5 sessions of data ................................................................................... 191 

Figure 4.19 Relative comparison of normalised average classification accuracies for LKM-

Self/ non-meditation EEG data when using the algorithm (CSP + STFT) for the three 

cases of 3 sessions, 4 sessions and 5 sessions of data ................................................ 191 

Figure 4.20 Comparison of average classification accuracies for LKM-Others/ non-meditation 

EEG data when using the algorithm CSP for the three cases of 3 sessions, 4 sessions 

and 5 sessions of data ................................................................................................. 193 

Figure 4.21 Relative comparison of normalised average classification accuracies for LKM-

Others/ non-meditation EEG data when using the algorithm CSP for the three cases of 

3 sessions, 4 sessions and 5 sessions of data .............................................................. 193 

Figure 4.22 Comparison of average classification accuracies for LKM-Others/ non-meditation 

EEG data when using the algorithm STFT for the three cases of 3 sessions, 4 sessions 

and 5 sessions of data ................................................................................................. 195 

Figure 4.23 Relative comparison of normalised average classification accuracies for LKM-

Others/ non-meditation EEG data when using the algorithm STFT for the three cases 

of 3 sessions, 4 sessions and 5 sessions of data .......................................................... 195 

Figure 4.24 Comparison of average classification accuracies for LKM-Others/ non-meditation 

EEG data when using the algorithm (CSP + STFT) for the three cases of 3 sessions, 4 

sessions and 5 sessions of data ................................................................................... 197 



xii 
 

Figure 4.25 Relative comparison of normalised average classification accuracies for LKM-

Others/ non-meditation EEG data when using the algorithm (CSP + STFT) for the 

three cases of 3 sessions, 4 sessions and 5 sessions of data ....................................... 197 

Figure 5.1 Workflow of inter-subject EEG classification across multiple meditation and non-

meditation sessions with different training–testing setups ......................................... 213 

Figure 5.2 Comparison of mean classification accuracies for LKM-Self/ non-meditation when 

using the three algorithms CSP, STFT or (CSP + STFT) for the cases of 3, 4 or 5 

session pairs of EEG data ........................................................................................... 220 

Figure 5.3 Comparison of mean classification accuracies for LKM-Others/ non-meditation 

when using the three algorithms CSP, STFT or (CSP + STFT) for the cases of 3, 4 or 5 

session pairs of EEG data ........................................................................................... 221 

Figure 5.4 Comparison of mean classification accuracies for LKM-Self/ non-meditation when 

using 3, 4 or 5 session pairs of EEG data for the three algorithms CSP, STFT or (CSP 

+ STFT) ...................................................................................................................... 221 

Figure 5.5 Comparison of mean classification accuracies for LKM-Others/ non-meditation 

when using 3, 4 or 5 session pairs of EEG data for the three algorithms CSP, STFT or 

(CSP + STFT) ............................................................................................................. 222 

Figure 5.6 Comparison of mean inter-subject (All Tests, Bottom 50%, Top 50%) and intra-

subject (previous study) classification accuracies for LKM-Self/ non-meditation when 

using 3 session pairs of EEG data for the three algorithms CSP, STFT or (CSP + 

STFT) ......................................................................................................................... 227 

Figure 5.7 Comparison of mean inter-subject (All Tests, Bottom 50%, Top 50%) and intra-

subject (previous study) classification accuracies for LKM- Others/ non-meditation 

when using 3 session pairs of EEG data for the three algorithms CSP, STFT or (CSP + 

STFT) ......................................................................................................................... 228 

Figure 5.8 Comparison of mean inter-subject (All Tests, Bottom 50%, Top 50%) and intra-

subject (previous study) classification accuracies for LKM-Self/ non-meditation when 

using 4 session pairs of EEG data for the three algorithms CSP, STFT or (CSP + 

STFT) ......................................................................................................................... 228 

Figure 5.9 Comparison of mean inter-subject (All Tests, Bottom 50%, Top 50%) and intra-

subject (previous study) classification accuracies for LKM- Others/ non-meditation 

when using 4 session pairs of EEG data for the three algorithms CSP, STFT or (CSP + 

STFT) ......................................................................................................................... 229 

Figure 5.10 Comparison of mean inter-subject (All Tests, Bottom 50%, Top 50%) and intra-

subject (previous study) classification accuracies for LKM-Self/ non-meditation when 

using 5 session pairs of EEG data for the three algorithms CSP, STFT or (CSP + 

STFT) ......................................................................................................................... 229 

Figure 5.11 Comparison of mean inter-subject (All Tests, Bottom 50%, Top 50%) and intra-

subject (previous study) classification accuracies for LKM- Others/ non-meditation 

when using 5 session pairs of EEG data for the three algorithms CSP, STFT or (CSP + 

STFT) ......................................................................................................................... 230 

  



xiii 
 

List of Tables 
 

Table 2.1 Summary of Participant and Meditation Details ...................................................... 71 

Table 2.2 Types of Meditation Techniques Used ...................................................................... 80 

Table 2.3 Summary of EEG Preprocessing .............................................................................. 81 

Table 2.4 Summary of Feature Extraction Methods Used ........................................................ 88 

Table 2.5 Summary of Performance of Pipelines and Classification Methods Used ............... 96 

Table 2.6 Summary of Study Objectives, Key Findings, and Limitations ............................. 100 

Table 2.7 Summary of EEG Frequency Bands Used .............................................................. 110 

Table 3.1 The five sessions selected for each of the 15 participants when in multiple session 

analysis ....................................................................................................................... 128 

Table 3.2 Average prediction accuracy (%) after conducting 25 tests each for 32 participants 

with single session using CSP and LDA (train 70%, test 30%) ................................. 131 

Table 3.3 Average prediction accuracy (%) after conducting 25 tests each for 15 participants 

with 5 sessions using CSP and LDA (train 70%, test 30%) ....................................... 132 

Table 3.4 Average prediction accuracy (%) after conducting 25 tests each for 15 participants 

with 5 sessions using CSP and LDA (train 20%, test 80%) ....................................... 133 

Table 4.1 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Self/ non-meditation EEG data using one of the three algorithms 

CSP, STFT or (CSP + STFT) per test for the case of 3 sessions of data .................... 168 

Table 4.2 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Others/ non-meditation EEG data using one of the three algorithms 

CSP, STFT or (CSP + STFT) per test for the case of 3 sessions of data .................... 170 

Table 4.3 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Self/ non-meditation EEG data using one of the three algorithms 

CSP, STFT or (CSP + STFT) per test for the case of 4 sessions of data .................... 172 

Table 4.4 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Others/ non-meditation EEG data using one of the three algorithms 

CSP, STFT or (CSP + STFT) per test for the case of 4 sessions of data .................... 174 

Table 4.5 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Self/ non-meditation EEG data using one of the three algorithms 

CSP, STFT or (CSP + STFT) per test for the case of 5 sessions of data .................... 176 

Table 4.6 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Others/ non-meditation EEG data using one of the three algorithms 

CSP, STFT or (CSP + STFT) per test for the case of 5 sessions of data .................... 178 

Table 4.7 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Self/ non-meditation EEG data using the algorithm CSP for the three 

cases of 3 sessions, 4 sessions and 5 sessions of data ................................................ 186 



xiv 
 

Table 4.8 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Self/ non-meditation EEG data using the algorithm STFT for the 

three cases of 3 sessions, 4 sessions and 5 sessions of data ....................................... 188 

Table 4.9 Average accuracy (%) calculated independently for each of the 12 participants when 

classifying LKM-Self/ non-meditation EEG data using the algorithm (CSP + STFT) 

for the three cases of 3 sessions, 4 sessions and 5 sessions of data ............................ 190 

Table 4.10 Average accuracy (%) calculated independently for each of the 12 participants 

when classifying LKM-Others/ non-meditation EEG data using the algorithm CSP for 

the three cases of 3 sessions, 4 sessions and 5 sessions of data ................................. 192 

Table 4.11 Average accuracy (%) calculated independently for each of the 12 participants 

when classifying LKM-Others/ non-meditation EEG data using the algorithm STFT 

for the three cases of 3 sessions, 4 sessions and 5 sessions of data ............................ 194 

Table 4.12 Average accuracy (%) calculated independently for each of the 12 participants 

when classifying LKM-Others/ non-meditation EEG data using the algorithm (CSP + 

STFT) for the three cases of 3 sessions, 4 sessions and 5 sessions of data ................ 196 

Table 5.1 Average accuracy (%) calculated independently for each of the 30 experiments 

conducted for classifying LKM-Self/ non-meditation EEG data using one of the three 

algorithms CSP, STFT or (CSP + STFT) per test for the case of 3 session pairs of data

 .................................................................................................................................... 215 

Table 5.2 Average accuracy (%) calculated independently for each of the 30 experiments 

conducted for classifying LKM-Others/ non-meditation EEG data using one of the 

three algorithms CSP, STFT or (CSP + STFT) per test for the case of 3 session pairs of 

data ............................................................................................................................. 216 

Table 5.3 Average accuracy (%) calculated independently for each of the 30 experiments 

conducted for classifying LKM-Self/ non-meditation EEG data using one of the three 

algorithms CSP, STFT or (CSP + STFT) per test for the case of 4 session pairs of data

 .................................................................................................................................... 217 

Table 5.4 Average accuracy (%) calculated independently for each of the 30 experiments 

conducted for classifying LKM-Others/ non-meditation EEG data using one of the 

three algorithms CSP, STFT or (CSP + STFT) per test for the case of 4 session pairs of 

data ............................................................................................................................. 217 

Table 5.5 Average accuracy (%) calculated independently for each of the 30 experiments 

conducted for classifying LKM-Self/ non-meditation EEG data using one of the three 

algorithms CSP, STFT or (CSP + STFT) per test for the case of 5 session pairs of data

 .................................................................................................................................... 218 

Table 5.6 Average accuracy (%) calculated independently for each of the 30 experiments 

conducted for classifying LKM-Others/ non-meditation EEG data using one of the 

three algorithms CSP, STFT or (CSP + STFT) per test for the case of 5 session pairs of 

data ............................................................................................................................. 219 

Table 5.7 Paired t-test results comparing classification accuracies across methods (d.f. = 5 for 

all comparisons) .......................................................................................................... 226 

Table 5.8 Hartigan’s Dip Test results for classification accuracy distributions across different 

session-pair conditions ............................................................................................... 233 



1 
 

Chapter 1 

Introduction 
 

1.1 Background and Motivation 

   

Meditation has gained considerable attention in scientific research [1, 2] due to its reported 

benefits for mental well-being [3, 4], emotional regulation [5, 6], and cognitive performance [7, 

8]. These outcomes have been increasingly supported by neurophysiological and behavioural 

studies, contributing to the growing integration of meditation practices into both clinical 

interventions and wellness programmes [9–11]. Among various forms of meditation, Loving-

Kindness Meditation (LKM) [12, 13] has shown particular promise for enhancing compassion, 

reducing stress, and cultivating positive affect. LKM involves the deliberate practice of 

generating and directing feelings of compassion, love, and goodwill, typically beginning with 

oneself and progressively extending to others, such as close acquaintances, neutral individuals, 

and eventually all beings. Unlike concentration-based practices that focus primarily on 

attentional control, LKM emphasises affective and emotional regulation, with the aim of 

cultivating prosocial emotions through repeated mental practice. In experimental settings, LKM 

is commonly implemented using guided instructions that structure the sequence and duration 

of these contemplative phases, making it suitable for controlled EEG-based investigation.  

To assess both long-term trait-level characteristics [14] and short-term meditative or non-

meditative mental states [15], researchers have traditionally relied on interviews and 

questionnaires as psychological assessment tools [16, 17]. However, with increasing interest in 

reliable and objective evaluation techniques, there has been a corresponding shift toward 

neuroimaging and computational approaches to quantify the effects of meditation [18, 19]. 



2 
 

Meditation is an umbrella term encompassing a wide range of contemplative practices; 

however, unless otherwise stated, the term “meditation” in this thesis is used specifically to 

refer to Loving-Kindness Meditation (LKM), which forms the exclusive focus of the 

experimental investigations. 

In meditation research, it is important to distinguish clearly between state and trait effects, as 

they reflect fundamentally different aspects of meditative practice [14, 15]. State meditation 

refers to transient, short-term changes in mental and neural activity that occur during or 

immediately following a meditation session, whereas trait meditation reflects longer-term, 

stable changes associated with sustained practice over months or years. State effects are 

typically examined by comparing meditation with non-meditation conditions, while trait effects 

are studied through longitudinal designs or comparisons between practitioners with differing 

levels of experience. This distinction is particularly relevant in EEG-based studies, as state-

level changes are more amenable to controlled experimental manipulation and classification, 

whereas trait-level effects are often confounded by individual variability. Accordingly, while 

both perspectives are acknowledged in this thesis, the experimental investigations primarily 

focus on state-level classification of meditation and non-meditation EEG, with multi-session 

analysis used to assess the stability of these state-related neural patterns over time. 

The terms meditation and mindfulness are often used interchangeably in the literature, despite 

representing partially overlapping but distinct constructs. Meditation broadly refers to a family 

of structured mental practices designed to cultivate specific cognitive, emotional, or attentional 

states, whereas mindfulness is commonly defined as a quality of present-moment awareness 

that may be cultivated through certain meditation practices but can also be conceptualised as a 

dispositional trait or everyday cognitive style. To avoid conceptual ambiguity and potential 

jingle-jangle fallacies, this thesis adopts a practice-based definition of meditation, focusing on 

clearly defined contemplative techniques rather than abstract psychological constructs. In line 



3 
 

with commonly used typologies of meditation research, which distinguish between attentional, 

constructive, and deconstructive practices [20, 21], Loving-Kindness Meditation is categorised 

as a constructive meditation practice, characterised by the deliberate cultivation of affective and 

prosocial mental states. Throughout this thesis, the term meditation therefore refers specifically 

to Loving-Kindness Meditation as an explicit, guided mental practice, rather than to 

mindfulness as a general cognitive disposition. 

Electroencephalography (EEG) has become a widely used technique for studying meditation 

due to its high temporal resolution, non-invasiveness, and portability [22, 23]. These features 

make EEG particularly suitable for capturing the subtle, time-sensitive changes in brain activity 

that occur during meditation [24, 25]. EEG enables the investigation of brain oscillations and 

connectivity patterns associated with different meditative and non-meditative mental states, 

offering insights into the neural mechanisms of mindfulness and concentration [26, 27]. In 

recent years, EEG has also been explored in Brain-Computer Interface (BCI) systems [28] 

aimed at mental state recognition, neurofeedback, and the development of adaptive meditation 

training tools [29]. These applications hold promise for personalizing meditation practices and 

may support long-term engagement through measurable progress tracking. 

Despite these advantages, reliably classifying meditative states using EEG remains a complex 

challenge. EEG signals are inherently non-stationary, noisy, and subject-specific, which makes 

it difficult to extract stable features that generalize across different users and recording sessions 

[30, 31]. Feature extraction and classification techniques must therefore be carefully selected 

and optimised to account for this variability [22]. While a range of machine learning and signal 

processing methods have been developed for general EEG analysis, their application to 

meditation data is often fragmented and lacks a cohesive computational framework [32, 33]. 

Many existing studies have focused on small datasets, often using single-session designs for 

between-group comparisons such as novice versus expert [29, 34], or collecting multiple 



4 
 

sessions primarily for limited pairwise analyses, such as meditation versus non-meditation [35, 

36] or before and after a meditation training programme [37, 38]. However, these studies rarely 

combine multiple sessions for comprehensive analysis, which limits the applicability of their 

findings to real-world scenarios where consistent performance across users and over extended 

time periods is essential. Real progress in meditation comes with continued practice across 

multiple sessions, making it especially valuable to understand meditation and non-meditation 

brain patterns across sessions, particularly when developing algorithms to support meditation 

progress. 

A limitation observed in the literature is the continued over-reliance on traditional frequency-

domain features, such as power spectral density measures, while more advanced spatial or time-

frequency methods remain underexplored [27, 39]. Although methods like Common Spatial 

Patterns (CSP) and Short-Time Fourier Transform (STFT) have demonstrated strong 

performance in other BCI contexts, particularly in motor imagery tasks [22, 32], they have not 

been systematically evaluated or combined for meditation-related EEG classification. 

Furthermore, the effect of training data size, particularly the number of EEG sessions, on 

classification accuracy is rarely studied, even though such factors are crucial for building 

reliable meditation monitoring systems. This gap becomes especially important when 

developing BCI pipelines intended for multi-session or subject-independent settings, which 

more closely reflect real-world deployment. A BCI pipeline typically refers to a structured 

sequence of computational steps, including data preprocessing, feature extraction, and 

classification, designed to interpret brain signals for specific tasks or feedback systems. 

The focus on Loving-Kindness Meditation in this thesis is deliberate and theoretically 

motivated. Compared to attentional practices such as focused attention or mindfulness-of-

breath, LKM places a stronger emphasis on affective and emotional regulation, making it 

particularly relevant for mental-health-oriented applications of EEG-based assessment. In 



5 
 

addition, LKM is typically implemented using structured, guided protocols that clearly 

delineate meditation and non-meditation periods, which is advantageous for controlled EEG 

experiments and supervised classification. From a practical perspective, the publicly available 

dataset used in this study was specifically designed around LKM and includes multiple sessions 

per participant, a feature that is rare in meditation EEG research and essential for investigating 

cross-session classification. While comparing multiple meditation types is an important 

direction for future research, restricting the scope of this thesis to a single, well-defined 

meditation practice allows for deeper methodological investigation of session-level variability 

and classification robustness without introducing additional confounding factors. 

This thesis is motivated by the need to bridge these gaps and move toward more generalizable 

and robust computational approaches for classifying Loving-Kindness Meditation (LKM) EEG 

data collected across multiple sessions and subjects. The research begins with a comprehensive 

computational review of the literature to identify dominant trends, overlooked techniques, and 

methodological weaknesses in existing EEG meditation studies. It then proceeds to achieve 

high single-session, intra-subject classification accuracies for LKM versus non-meditation and 

explores recurring neural patterns across multiple sessions for each individual, laying the 

foundation for robust personalized models. It subsequently develops and evaluates novel BCI 

pipelines using CSP, STFT, and their fusion, focusing on classifying LKM versus non-

meditation states. These pipelines are assessed under both intra-subject and inter-subject 

conditions using multiple-session EEG data, addressing the variability across individuals and 

recording sessions that is often ignored in simpler models. 

Although the core of this thesis lies in experimental development and evaluation, it also yields 

several theoretical and methodological insights relevant to EEG-based meditation research. 

These include the influence of training session count on classification accuracy, the viability 

and benefits of combining spatial and time-frequency feature extraction techniques, and a 



6 
 

comparative understanding of model performance in both personalized (intra-subject) and 

generalised (inter-subject) settings. All experimental investigations are conducted using a 

publicly available EEG dataset specifically collected for meditation research, ensuring 

consistency, transparency, and reproducibility. This dataset comprises clearly labelled, 

multiple-session EEG recordings covering Loving-Kindness Meditation and non-meditative 

control states, thereby providing a reliable foundation for the comparative analysis of different 

classification pipelines. The EEG dataset used in this study is publicly available at 

https://doi.org/10.18112/openneuro.ds003816.v1.0.1 and was obtained from OpenNeuro. 

The scope of this thesis is clearly defined: it focuses solely on EEG-based analysis and does not 

incorporate other modalities such as fNIRS or hybrid systems. By working within the 

constraints of a single, well-structured dataset and avoiding the variability introduced by multi-

modal integration, the study maintains a high degree of experimental control while still 

addressing real-world challenges such as session-to-session variability and subject diversity. 

These design decisions reflect a deliberate balance between methodological precision and 

practical applicability, advancing the broader objective of developing scalable, generalizable 

EEG-based meditation classification systems with future potential in mental health support, 

neuroadaptive technology, and personalized wellness applications. 

 

1.2 Problem Statement 

 

The integration of electroencephalography (EEG) with computational algorithms such as signal 

processing and machine learning has shown significant promise for assessing meditation states, 

yet several critical gaps remain. Firstly, there is an absence of a comprehensive computational 

review that systematically categorizes and evaluates existing feature extraction and 



7 
 

classification techniques specifically for meditation-related EEG studies, which limits clarity 

on the current research landscape. Secondly, while intra-subject and inter-subject (subject-

independent) EEG classification during meditation has been explored, most existing studies 

rely on pairwise session comparisons and overlook the temporal variability and consistency of 

neural patterns across multiple sessions. Thirdly, the generalizability of classification models 

across individuals (subject-independent classification) remains underexplored and challenging, 

especially in the context of multi-session meditation EEG data that tends to be highly 

individualized. Moreover, although numerous feature extraction techniques exist, few studies 

provide comparative evaluations of selected high-performing methods along with their fusion, 

such as the combination of spatial and time-frequency features, in both intra-subject and inter-

subject multi-session EEG classification scenarios related to meditation. Finally, the impact of 

varying the number of training sessions on classification performance using multiple-session 

meditation EEG data is not well understood, which limits the practical deployment of such 

models in real-world neurofeedback or meditation monitoring applications. 

This thesis addresses these issues through a combination of computational review and 

experimental evaluations, with a focus on designing and benchmarking EEG classification 

pipelines for meditation state detection across single and multiple sessions, and across both 

intra- and inter-subject contexts. 

 

1.3 Research Objectives 

 

The research objectives of this thesis are outlined below: 



8 
 

1. To conduct a comprehensive computational review of EEG-based feature extraction and 

classification techniques for meditation assessment to address a notable gap in the 

literature where such a review has not previously been presented. 

2. To classify single-session, intra-subject meditation versus non-meditation EEG data 

with high accuracy using appropriate feature extraction and classification techniques, 

and to use the results to identify neural patterns associated with meditation and non-

meditation mental tasks. 

3. To investigate the presence of consistent neural patterns across multiple sessions within 

subjects for both meditation and non-meditation tasks using intra-subject classification. 

4. To develop and evaluate BCI pipelines incorporating various feature extraction 

algorithms for intra-subject, multiple-session EEG classification to obtain optimal 

accuracies, where a new session pair is classified based on training from previous 

sessions. Furthermore, to investigate how the fusion of such successful feature 

extraction algorithms impacts classification performance. 

5. To study how varying the number of training session pairs influences the classification 

accuracy of intra-subject, multiple-session meditation and non-meditation EEG data 

using the developed BCI pipelines. 

6. To implement and evaluate subject-independent, multiple-session classification 

pipelines under conditions consistent with intra-subject classification, ensuring training 

and testing are conducted across different individuals. 

7. To compare the performance of intra-subject and inter-subject multiple-session 

classification and investigate how changing the number of training session pairs affects 

classification accuracy in the subject-independent context. 



9 
 

 

1.4 Thesis Organization 

 

This thesis is organised into six chapters, each addressing a distinct part of the research process: 

• Chapter 1: Introduction 

Introduces the background and motivation for the study, presents the problem statement, 

outlines the research objectives, and provides an overview of the thesis structure. It 

briefly touches on the methodological approach and scope of the study. 

 

• Chapter 2: A Computational Review of Feature Extraction and Classification 

Techniques for Meditation Assessment Using EEG 

Presents a systematic review of existing literature on EEG-based feature extraction and 

classification methods used for meditation assessment. It identifies methodological 

trends, summarizes commonly used techniques, highlights current gaps in the literature, 

and offers valuable insights and guidance for future research in the field of meditation 

EEG. 

 

• Chapter 3: Common Spatial Pattern for Classification of Loving Kindness 

Meditation EEG for Single and Multiple Sessions 

Investigates intra-subject classification of meditation and non-meditation EEG data 

using the Common Spatial Pattern (CSP) algorithm. It examines both single-session and 

multiple-session scenarios and evaluates the consistency of neural patterns across 

sessions within individuals. 



10 
 

 

• Chapter 4: Novel Machine Learning-Driven Comparative Analysis of CSP, STFT, 

and CSP-STFT Fusion for EEG Data Classification Across Multiple Meditation 

and Non-Meditation Sessions in BCI Pipeline 

Develops and evaluates BCI pipelines using CSP, Short-Time Fourier Transform 

(STFT), and a fusion of both for classifying intra-subject, multiple-session EEG data. 

This chapter also investigates how varying the number of training sessions affects 

classification accuracy and explores the advantages of feature fusion. 

 

• Chapter 5: Machine Learning-Based Comparative Analysis of Subject-

Independent EEG Data Classification Across Multiple Meditation and Non-

Meditation Sessions 

Extends the classification framework to subject-independent scenarios, applying the 

same three BCI pipelines. It evaluates classification performance across individuals, 

compares results with intra-subject outcomes, and examines the effect of training 

session size on generalizability. 

 

• Chapter 6: Conclusion and Future Work 

Summarizes the research findings and key contributions across all chapters. It discusses 

the practical implications of the results, highlights limitations encountered during the 

study, and proposes future research directions to enhance EEG-based meditation state 

detection in real-world applications. 

 



11 
 

  



12 
 

Chapter 2 

A Computational Review of Feature Extraction and 

Classification Techniques for EEG-Based Meditation 

Assessment 
 

Abstract 

Meditation research using electroencephalography (EEG) has gained significant attention, with 

computational techniques playing a key role in feature extraction and classification. This review 

examines EEG-based meditation assessment by systematically analyzing 164 peer-reviewed 

studies. The selected studies were categorized based on meditation study characteristics, EEG 

preprocessing techniques, feature extraction methods, classification approaches, and EEG 

frequency bands utilised. Feature extraction methods were grouped into ten major categories 

based on their computational approaches, underlying mathematical principles, and relevance to 

EEG signal analysis in meditation studies. Frequency domain analysis contributed the most, 

accounting for 62.0% of extracted features. Classification techniques in 40 articles were 

analysed based on the number of methods used, grouping studies into those employing one, 

two, or three or more techniques. Support vector machines (55.0%) and artificial neural 

networks (47.5%) were among the most frequently used techniques, with some studies applying 

multiple methods. Alpha (90.2%) was the most frequently examined EEG frequency band in 

meditation studies, followed by Theta (77.4%). The findings highlight key trends, variations, 

and challenges in data acquisition, preprocessing, feature extraction, and classification. Due to 

methodological differences across studies, direct comparisons of classification accuracies were 

not possible. The review provides a structured synthesis of computational methods used in 

meditation EEG studies, identifies key research gaps, and discusses future directions, 



13 
 

emphasizing the need for standardized methodologies to enhance reproducibility and 

comparability. 

 

Keywords: Feature Extraction, Classification, EEG (Electroencephalography), Meditation, BCI 

(Brain Computer Interface), Preprocessing 

 

2.1 Introduction 

 

2.1.1. Background on Meditation and EEG Neuroimaging 

 

Meditation [40, 41], although practised for thousands of years, has gained significant popularity 

in modern stressful lifestyle due to scientific evidence supporting its ability to enhance self-

awareness, reduce stress, and improve overall mental health [4, 5, 11]. Although there are many 

different meditation techniques that may differ in how they are practised, studies have been 

conducted on many of these techniques, demonstrating various similar benefits [1, 3, 7]. Studies 

show that there are short-term and long-term effects of meditation practice and referred to as 

‘State’ and ‘Trait’ changes [15, 42]. 

Scientific methods used to study meditation include neurophysiological methods such as 

Electroencephalography (EEG) [43–46], Functional Magnetic Resonance Imaging (fMRI) [47], 

and Functional Near-Infrared Spectroscopy (fNIRS) [48]; physiological methods such as Heart 

Rate Variability [9], Cortisol Levels [49], and Blood Pressure; and psychological and 

behavioural methods such as Mindfulness and Meditation Scales, Questionnaires, and 

Interviews [16, 17, 50]. With the advancement of mobile phones and electronic technology, we 

have witnessed the development of meditation-supporting devices and mobile apps [18, 51, 52], 



14 
 

which have, in turn, increased the interest of both scientists and the general public in meditation 

[6, 53, 54]. 

EEG is a non-invasive neuroimaging technique that records electrical activity generated by 

neuronal oscillations in the brain [55, 56], collected through electrodes placed on the scalp [57]. 

EEG data collected over a certain duration consists of oscillatory signals over time. In most 

studies, this data is converted into the frequency domain, which is divided into five bands 

associated with specific cognitive and physiological states. These bands are Delta (0.5–4 Hz), 

Theta (4–8 Hz), Alpha (8–13 Hz), Beta (13–30 Hz), and Gamma (>30 Hz) [58–60]. The strength 

of these bands varies with different mental tasks, such as deep sleep, light sleep, meditation, 

wakeful rest, active thinking, and advanced problem-solving [56, 61]. 

Research on EEG in general has significantly flourished in the recent past due to improvements 

in hardware and software resources and reduction in prices. This includes improvements in 

EEG sensors and the processing power of devices. At the same time, with the development of 

feature extraction and classification algorithms[62–64], along with improvements in 

preprocessing and EEG data cleaning methods [30, 31, 65], there has been a significant 

improvement in the results obtained from raw EEG data [66–68]. 

EEG has been successfully used in various studies related to meditation, where different 

patterns in EEG data are mapped to various mental tasks associated with meditation. Some 

examples of such studies include comparisons between meditation and non-meditation [69, 70], 

novice and expert meditators [71, 72], different types of meditation [13, 73], various wavebands 

[74, 75], state and trait changes [26], and the effects of meditation training on novice meditators 

[76], among others. 

Although studies on meditation EEG date back over 50 years [19, 77–79], research in this area 

is significantly less extensive compared to fields like motor imagery [28]. When examining 



15 
 

motor imagery EEG, previous research reveals a wealth of studies on computational methods, 

including feature extraction and classification [80–82]. The extensive body of research in motor 

imagery is well-documented in several review articles [22, 32]. In contrast, there are far fewer 

studies on computational methods, such as feature extraction and classification, in the context 

of meditation EEG. Furthermore, there are no systematic reviews that have been conducted to 

specifically examine these computational methods in meditation EEG, other than one instance 

of a review paper that lightly touches on machine learning methods used in meditation studies 

[2]. In this review, we explore how feature extraction and classification techniques have been 

effectively applied in various studies to analyse meditation EEG data, addressing a critical 

knowledge gap. 

 

2.1.2. Overview Objectives of the Review 

 

There is a significant body of research, including numerous original studies and review articles, 

published on meditation EEG. While many of these studies employ a variety of feature 

extraction and classification methods, existing review articles rarely focus specifically on the 

computational techniques involved in these processes. Recognizing this gap, the present review 

aims to address it. We selected peer-reviewed meditation EEG research articles published over 

the past 15 years, enabling us to trace the evolution of computational methods in feature 

extraction and classification. Additionally, this study captures important complementary aspects 

such as participant and meditation characteristics, data cleaning and preprocessing strategies, 

and the usage of frequency bands in meditation EEG research. Although this review maps a 

broad range of computational techniques used in meditation EEG research, a systematic 

literature review methodology was adopted to ensure transparency, replicability, and objective 

study selection. Unlike scoping reviews, which are primarily exploratory in nature, this work 



16 
 

applies predefined research questions, inclusion criteria, and structured synthesis to provide a 

reproducible computational reference for the field. 

 

2.2 Methodology 

 

2.2.1. Overview of the Systematic Literature Review Approach 

 

A systematic literature review approach was used in this review to uphold the quality and 

reproducibility of our work in a transparent manner. The established guidelines in PRISMA 

were utilised for identifying, screening, and selecting research articles relevant to meditation 

EEG feature extraction and classification. This rigorous approach allowed us to capture a wide 

spectrum of studies and provided a comprehensive basis for our subsequent analysis. 

 

2.2.2. Research Questions (RQ) 

 

The review was built around the following research questions: 

RQ1: What feature extraction methods have been used in EEG for meditation 

assessment? 

RQ2: What classification techniques have been applied in meditation EEG studies, and 

how have their performances, study objectives, key findings, and limitations been evaluated in 

those studies? 

RQ3: How have preprocessing and data cleaning techniques been used in studies that 

apply feature extraction and/ or classification methods in meditation EEG? 



17 
 

RQ4: What types of meditation techniques, participants, and EEG frequency bands have 

been used in studies that apply feature extraction and/ or classification methods in meditation 

EEG? 

 

2.2.3. Search Strategy 

 

A comprehensive search was conducted across multiple scientific databases, including Scopus, 

PubMed, SpringerLink, ScienceDirect, and IEEE Xplore. We used two sets of keywords along 

with Boolean operations in the search: “Meditation AND EEG AND Feature Extraction” and 

“Meditation AND EEG AND Classification.” The keyword searches were applied to the full 

searchable fields of the databases (including titles, abstracts, and indexed content), rather than 

being restricted to titles or abstracts alone, allowing studies primarily indexed under related 

terms such as mindfulness-based practices (e.g., MBSR or Vipassana) to be captured. 

This review includes studies published between 2010 and 2024, covering a 15-year period to 

capture the evolution of computational techniques in meditation EEG analysis. This timeframe 

ensures the inclusion of recent advancements while providing sufficient historical context for 

understanding methodological trends.  

We also applied filters to restrict the search to peer-reviewed articles published in English. This 

approach ensured that our search was both comprehensive enough to capture all relevant studies 

and selective enough to maintain quality and relevance. 

 

 

 



18 
 

2.2.4. Inclusion and Exclusion Criteria 

 

To determine which studies to include in the review, we defined clear inclusion and exclusion 

criteria. The inclusion criteria were as follows: studies that focused on EEG-based meditation 

assessment, employed computational methods for feature extraction and/or classification, and 

provided sufficient methodological details. The exclusion criteria included: studies that did not 

involve EEG data, were not peer-reviewed, or focused solely on subjective measures without 

any computational analysis. This systematic filtering process ensured that only studies relevant 

to our review’s focus were included. 

 

2.2.5. Screening and Selection Process 

 

The initial search using keywords and relevance yielded a large number of article titles along 

with their abstracts, from which duplicates and non-English articles were removed. Then, the 

selected article titles and abstracts were screened, and only those relevant to the review were 

retained. Next, the full-text versions of the selected studies were retrieved. After that, the 

remaining articles underwent a detailed full-text screening based on our inclusion and exclusion 

criteria. Finally, 164 research articles were selected for inclusion in this review. 

A PRISMA flow diagram was used to illustrate the number of records at each stage: 

identification, screening, and final inclusion, ensuring a transparent account of the selection 

process. This is presented in Figure 2.1. 

 



19 
 

 

Figure 2.1 PRISMA flow diagram for selecting articles for the systematic review 

 

2.2.6. Data Extraction and Categorization 

 

For each selected study, we extracted key details such as participant characteristics, meditation 

characteristics (including techniques), EEG frequency bands, EEG preprocessing methods, and 



20 
 

EEG feature extraction and classification techniques. From studies that applied classification 

techniques, additional details, including the performance of pipelines, study objectives, key 

findings, and study limitations, were also extracted. 

The extracted data were organised into structured tables to facilitate analysis: 

Table 2.1: Summary of Participant and Meditation Details 

Table 2.2: Types of Meditation Techniques Used 

Table 2.3: Summary of EEG Preprocessing 

Table 2.4: Summary of the Feature Extraction Methods Used 

Table 2.5: Summary of the Performance of Pipelines and Classification Methods Used 

This systematic data extraction allowed us to categorize the diverse methods into coherent 

groups for detailed analysis. 

 

The summarized results are further analysed in the following sections of the review: 

"Meditation Study Characteristics and Preprocessing" – Discusses participant details, 

meditation types, comparison group details, EEG preprocessing, and noise handling methods. 

"Feature Extraction Techniques for Meditation EEG Assessment" – Reviews various 

feature extraction methods used in meditation EEG research, categorized based on 

computational approaches. 

"Classification Techniques for Meditation EEG Assessment" – Explores machine 

learning and other classification techniques applied to meditation EEG data, highlighting 

pipeline performances. 



21 
 

By structuring the extracted data within these sections, this review provides a comprehensive 

analysis of computational methods used in meditation EEG research. 

 

2.2.7. Data Synthesis and Analysis 

 

After categorizing the extracted data, a structured synthesis was conducted to identify trends, 

gaps, and key insights in the use of feature extraction and classification techniques for 

meditation EEG assessment. This synthesis involved both qualitative and quantitative analyses, 

comparing study methodologies, extracted features, classification techniques, and reported 

outcomes. 

Key findings were summarized through the following: 

Table 2.6: Summary of Study Objectives, Key Findings, and Limitations – Provides a 

high-level overview of the main research questions addressed, significant findings, and reported 

limitations for studies that applied classification techniques. 

Table 2.7: Summary of EEG Frequency Bands Used – Categorizes EEG frequency 

bands analysed in meditation studies, highlighting those most frequently associated with 

meditation. 

 

The synthesized insights are further discussed in the "Conclusion" section under two subtopics: 

"Recap of the Review’s Major Insights" – Summarizes patterns in feature extraction and 

classification techniques across studies. 

"Mapping Findings to Current and Future Meditation EEG Studies" – Connects the 

review’s findings to broader trends in EEG-based neuroimaging and future research directions. 



22 
 

By structuring the synthesis through these tables and discussions, the review presents a 

comprehensive analysis of computational techniques in meditation EEG research, highlighting 

both established methodologies and areas requiring further exploration. 

 

2.2.8. Limitations of the Methodology 

 

Finally, we acknowledge certain limitations in our review process. These include potential 

publication bias, variations in EEG data collection and preprocessing methods, and 

inconsistencies in the reporting of computational techniques. Additionally, the exclusion of non-

English studies might limit the comprehensiveness of our review. 

Regarding EEG data collection, significant differences exist among studies due to variations in 

devices, recording protocols, and participant-related factors. These include the types of 

meditation techniques used, number of participants, experience levels, comparison groups, and 

whether the study examined state and/or trait characteristics. Such differences inherently affect 

the EEG data collected and subsequently influence the reported performance metrics. 

Furthermore, differences in preprocessing methods significantly impact results. These include 

the type of filtering methods applied, whether automated cleaning techniques were used, and 

whether manual cleaning with visual inspection was performed. Variations in these approaches 

can lead to inconsistencies in final outcomes. 

Since our primary focus is on feature extraction and classification methods in EEG-based 

meditation studies, the reported performance metrics and classification accuracies were 

considered as given in the original studies. However, due to the substantial methodological 

differences between studies, we were unable to directly compare classification accuracies 

across studies. This remains the most significant limitation of our review. 



23 
 

 

2.3 Meditation Study Characteristics and Preprocessing 

 

2.3.1. Meditation Study Characteristics 

 

When reviewing research articles on feature extraction and classification of meditation EEG 

data, several dataset characteristics were found to substantially influence study outcomes. These 

include the type of meditation technique employed, whether the study examined state or trait 

effects, participant experience level (novice, intermediate, or expert), total sample size, and 

whether intra-subject or inter-subject analyses were performed. Such design choices directly 

affect EEG patterns, feature selection strategies, and reported classification performance. 

Therefore, as part of this literature review, detailed information related to meditation protocols 

and participant characteristics was systematically extracted from each article. Rather than 

providing a generic overview, this analysis highlights how meditation type, participant 

experience, and experimental design choices have been handled across prior studies, enabling 

more informed methodological decisions when designing new experiments. 

The summary of meditation and participant details collected from 164 research articles is shown 

in Table 2.1 and Table 2.2. In Table 2.1, each line represents the details of a single research 

article along with the reference to the article. The first piece of information included in Table 

2.1, in the column labelled ‘Meditation Type,’ is the type of meditation technique used in each 

of the studies. As mentioned in the introduction, there are many types of meditation techniques, 

and a study in a research article may use one or several meditation techniques depending on the 

study's requirements. It should be noted that meditation practices are inherently diverse, and 

attempts to categorise them into a small number of discrete styles are often approximate. Many 

techniques share common elements, and hybrid or overlapping practices are frequently 



24 
 

observed in both traditional and contemporary meditation research. The next piece of 

information collected is the state/trait characteristic of meditation used in each of the research 

articles. As explained in the introduction, state/trait characteristics refer to short-term/long-term 

changes occurring in participants practicing meditation. In feature extraction and classification, 

we aim to identify different groups related to meditation studies. Here, in most cases, the 

differences among these groups are related to the state/trait influence of meditation. Therefore, 

the state/trait characteristic related to each article is included in Table 2.1, under the column 

title ‘State/Trait’. 

In meditation EEG data collection studies, the meditation experience of the participant is also 

a significant factor that contributes to the results. Therefore, for each article, we identified the 

participants' meditation experience level and included it in Table 2.1 under the column title 

‘Experience Level’. When going through the articles, we identified three main groups of 

participants based on meditation experience level, which can be labelled as novice (or control), 

intermediate, and expert. Depending on the requirements of each study, we observed the use of 

one or more experience levels. At the same time, the number of participants involved in each 

study is also an important piece of information, and this is also included in Table 2.1 under the 

column title ‘Total No of Participants’. Here, when a research article focuses on more than one 

experience level, we noted the number of participants for each experience level used in the 

study. 

Another important point in this type of research is whether the study conducts intra-subject, 

inter-subject, or both types of analyses. In intra-subject analysis, different instances of the same 

person are compared, while in inter-subject analysis, certain characteristics are compared 

among different participants. At the same time, a few instances were noticed where both 

comparisons were conducted in the same research article. Therefore, for each article, this 

information was also collected and included in Table 2.1 under the column title ‘Intra/Inter-



25 
 

Subject Analysis’. As the final information under Table 2.1, we have included the comparison 

groups used in each of these studies under the column title ‘Comparison Groups’. This 

information is included in Table 2.1 to explicitly document how prior studies structured their 

comparison groups and analytical scope, allowing readers to identify dominant experimental 

patterns and methodological gaps in the existing literature. 

Table 2.2 summarizes the types of meditation techniques used in the 164 research articles we 

studied, identifying 39 different techniques. Of these, 152 articles explicitly stated the 

meditation techniques used, while the remaining 12 were excluded from this table. While many 

papers demonstrate the use of a single meditation technique, a few articles collected EEG data 

from participants practicing different meditation types, with a selected number of participants 

for each. Here we can see that, among 152 research articles, 37 articles have used Mindfulness 

Meditation in meditation EEG feature extraction and classification studies, which is 24.3%. 

Next, we observed Focused Attention Meditation and Vipassana Meditation, both techniques 

used in 14 research articles, which accounts for 9.2%. Out of the 39 meditation techniques used, 

21 of them were used only once or twice among the 152 articles. In Table 2.2, the percentage 

represents how often each meditation technique was used per 100 papers. Since some research 

articles employed multiple techniques, the sum of the percentages exceeds 100. 

 

2.3.2. Preprocessing 

 

Preprocessing done in a well-planned manner is an important step in EEG pipeline development 

that significantly affects the performance of the pipeline. Therefore, in the research articles 

selected to study meditation EEG feature extraction and classification, certain characteristics 

related to preprocessing were also studied. The results of this study provide a consolidated view 



26 
 

of preprocessing practices used in meditation EEG research, revealing common design choices, 

inconsistencies, and underreported steps that directly influence reproducibility and pipeline 

performance. Under EEG preprocessing, the two main tasks are removing noise and artifacts 

from the raw EEG data and making modifications to the raw EEG data, such as splitting and 

filtering. This would result in a slightly modified EEG dataset that is much cleaner and easier 

to handle in the subsequent steps of the pipeline. Table 2.3 provides a summarized set of details 

on preprocessing related to the 164 research articles we reviewed. 

In Table 2.3, the first piece of information for each article is the number of channels used in 

each study. The number of channels used to collect the EEG data plays a significant role in a 

study because the more electrodes used, the better the overall quality of the data collected. 

When going through the selected articles, we can see that some studies analyse the entire EEG 

dataset, while others break the EEG data into small segments called epochs, with analysis 

conducted independently on each epoch. Additionally, some studies that use epochs go a step 

further by using overlapping epochs. Overlapping epochs are used to minimize the effect of 

signal loss at the splitting points between epochs. In this case, the lost data will be included in 

the next adjacent epoch due to the overlapping. In Table 2.3, under the column 'Epoch Size 

(Overlap)', we have included both the epoch size and overlap size for each research article that 

used them in their studies. 

Signal filtering methods are used to improve the quality and reduce the noise in EEG data. In 

addition to commonly discussed physiological and environmental noise, EEG recordings may 

also be affected by artefacts such as electrode–skin impedance variations, subject motion, 

amplifier saturation, and digitisation noise. The idea is to remove data related to certain 

frequency ranges that are not relevant to the study while retaining data from the frequency 

ranges that are important. Filtering therefore represents one component of a broader set of 

preprocessing strategies, which may be combined with other artefact-handling approaches 



27 
 

depending on data quality and study objectives. Additionally, filtering can be used to remove 

power line interference occurring at 50/60 Hz. If any filtering methods were applied in the 

research articles we reviewed, they are listed in the column 'Signal Filtering Methods' in Table 

2.3. 

When considering noise and artifact removal from EEG, this can be done either by using 

automated methods and algorithms or by observing the data graphically and manually 

identifying and removing artifacts. While reviewing past research articles, we observed 

instances where neither method was used, instances where one method was used, and instances 

where both methods were employed. In Table 2.3, the column 'Automated Artifact Removal' 

lists the methods used in a manuscript if automated cleaning was performed, while the column 

'Manual Artifact Removal' indicates ‘Yes’ or ‘Not Mentioned’ to show whether manual cleaning 

was conducted in each of the research articles. 

 

2.4 Feature Extraction Techniques for Meditation EEG Assessment 

 

2.4.1. Introduction of Feature Extraction Techniques 

 

In the study of meditation EEG, feature extraction is a critical section where crucial information 

is extracted for research purposes. Given the broad scope of this topic, we have categorized 

these features into ten groups for clarity. These categories are: Frequency Domain Analysis, 

Time-Frequency and Spectral Entropy Methods, Amplitude Features, Dimensionality 

Reduction and Decomposition, Phase and Synchrony Analysis, Entropy-Based Measures, 

Connectivity and Coherence Measures, Spatial, Microstate, Network and Topological Features, 

Statistical Features, and Advanced Modelling, Machine Learning, and Complex Features. In 



28 
 

the following sections, we will demonstrate how these different groups of features have been 

used efficiently in past literature to extract valuable information, while identifying various 

feature types for each of these feature categories. 

 

2.4.2. Frequency Domain Analysis 

 

As the first category in meditation EEG feature extraction, we will examine “Frequency 

Domain Analysis”. Here, the EEG signal is converted into the frequency domain from its 

original time domain, allowing researchers to study the power distribution across different 

frequencies. In brain signal analysis, this frequency spectrum is typically divided into distinct 

bands: Delta (0.5-4 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz), and Gamma (>30 

Hz). Changes in the power levels of these frequency bands are often associated with various 

mental tasks.  

When we review research conducted on EEG, we can observe that a significant amount of work 

has used spectral analysis to understand oscillatory patterns in the participants. Meditation EEG 

research also follows this trend, and we can clearly observe many instances where Spectral 

Analysis [8, 35, 83–90] has been utilised in relevant research articles. Among the many methods 

used under this, the Power Spectral Density (PSD) [23, 29, 33, 36, 37, 90–122] is one of the 

most commonly used methods. The PSD represents the distribution of power across different 

frequencies in the EEG signal. It works by breaking the signal into frequency bands and 

quantifying the power (or energy) within each band. The power variations in these frequency 

bands can then be used as features to compare different mental states or meditation tasks. The 

Bartlett method [33] is an improved technique used to calculate the PSD of an EEG signal. It 

divides the signal into smaller, non-overlapping segments, calculates the PSD for each segment, 



29 
 

and then averages the resulting periodograms to obtain a smoother estimate with reduced 

variance. 

Among the methods used to calculate the PSD, the Fast Fourier Transform (FFT) [12, 39, 93, 

102, 103, 109–111, 121, 123–142] is the most commonly used technique in EEG feature 

extraction. FFT transforms EEG signals from the time domain into the frequency domain, 

allowing us to observe the strength (power) of different frequencies. At the same time, the 

Blackman-Harris window [33] is a tool that is used to enhance accuracy when extracting 

frequency strengths from EEG data. It applies smooth edges to the signal, reducing spectral 

leakage and errors when calculating the PSD using FFT. The Discrete Fourier Transform (DFT) 

[143], on the other hand, is used less with meditation EEG due to its higher computational 

complexity and lower memory efficiency. DFT provides the basic mathematical foundation of 

Fourier Transform and can produce reliable PSD outputs when dealing with smaller datasets 

without real-time requirements. 

The Spectral Power Ratio (SPR) [92, 93, 105, 111, 137, 144–149] is a feature set derived from 

the PSD that can be used in classifying various meditation-related mental tasks. The SPR 

measures the power ratio between two frequency bands in EEG, and among the five bands 

mentioned, any two can be selected for this calculation, such as the Alpha/Beta Ratio, 

Theta/Beta Ratio, Delta/Theta Ratio, etc. Welch’s Modified Periodogram [25, 36, 84, 103, 108, 

137, 150, 151] is an improved method of calculating the PSD. The basic idea is that the EEG 

signal is segmented into small, often overlapping, windows, and a windowing function is 

applied before computing the periodogram for each segment using a method such as FFT. Then, 

the results of all windows are averaged to obtain a smooth and reliable PSD estimate, providing 

a clean feature set for further studies. Compared to Bartlett’s method, Welch’s approach further 

reduces variance while providing improved spectral smoothness through overlapping and 

windowing. 



30 
 

In some cases, specific sections of the PSD are used as features in meditation EEG analysis. 

PSD is divided into five bands, and you may refer to the features by the name of the band when 

you use a particular band’s information as features. Also, there are instances where information 

from several bands is independently used in meditation EEG studies, and we can generally label 

this as Band Power [12, 25, 35, 96, 116, 140, 144, 147, 148, 151–162]. Instances can be 

observed where characteristics of individual frequency bands are used as features when 

studying and classifying various meditation-related EEG data. These instances can be referred 

to by the band names, and some recent research examples where the power of each band is used 

are Delta [163], Theta [134, 163, 164], Alpha [14, 27, 38, 98, 102, 117, 120, 127, 131, 134, 137, 

139, 163, 165–169], Beta [133, 134, 163], and Gamma [134, 163, 170]. 

In spectral analysis, it is sometimes valuable to identify the frequency with the highest power, 

as this can help in understanding different mental states. In meditation EEG feature extraction 

studies, a few instances have been observed where this peak frequency was used, with studies 

employing methods such as the Frequency of the Main Peak [127, 165] and Instantaneous Peak 

Frequency Calculation [145].  

Along with EEG spectral feature extraction, we observed Hanning Windowing [39, 103] as a 

helpful method for improving the quality of frequency-related features by reducing spectral 

leakage. In the process of converting EEG data into the frequency domain, the data is segmented 

into small-sized windows, which can create artificial frequencies at the edges of those windows, 

known as spectral leakage. The Hanning window minimizes this problem by smoothly tapering 

the edges of the window toward zero, thereby improving the accuracy of the obtained 

frequencies. 

 



31 
 

2.4.3. Time-Frequency and Spectral Entropy Methods 

 

As the second category in meditation EEG feature extraction, we will examine “Time-

Frequency and Spectral Entropy Methods”. Both these techniques are used to study how brain 

patterns change over time. Here, the Time-Frequency Distribution (TFD) shows how the power 

of each frequency changes over time, while the Spectral Entropy, which is derived from the 

Power Spectral Density (PSD), quantifies the signal randomness, allowing for the study of 

dynamic brain states in EEG data. TFD enables the examination of how a mental task, such as 

meditation, affects certain frequency bands over time and has been widely used in meditation 

EEG research [23, 38, 104, 118, 164, 171–174]. For example, when the mind relaxes during 

meditation, changes in the Alpha or Theta band powers may be observed, while an increase in 

the Gamma band might be associated with tasks involving heightened cognitive processing or 

problem-solving. Several methods are commonly used to obtain TFD in meditation EEG 

research, including Short-Time Fourier Transform, Discrete Wavelet Transform, Wavelet 

Transform, and Stockwell Transform. 

The Short-Time Fourier Transform (STFT) [100, 118, 146, 175] is a method used to study how 

the strength of frequencies in EEG data changes over time. STFT breaks the data into small 

segments called windows (often overlapping) and applies the Fourier Transform on each 

window to get a frequency spectrum. STFT is capable of analyzing transient brain activities, 

such as those occurring during meditation, and supports Brain-Computer Interface (BCI) 

applications that require real-time EEG processing. 

The Discrete Wavelet Transform (DWT) [89, 107, 128, 130, 148, 156, 165, 176–181] is another 

commonly used method for obtaining TFD patterns in meditation EEG data. DWT first breaks 

the signal into two components using low and high-pass filters to extract approximation and 

detail coefficients. These results are then downsampled to reduce the data size while retaining 



32 
 

the essential features of the signal. This process is repeated for multiple levels of decomposition 

until a clean time-frequency spectrum is obtained, especially for non-stationary signals like 

EEG. DWT is effective in reducing noise in the signal while preserving the important features.  

The Continuous Wavelet Transform (CWT) [182, 183] uses a continuous range of scales, unlike 

the discrete set of scales used in DWT, when calculating the TFD for EEG. CWT provides a 

highly detailed decomposition of the EEG data, offering more precise time-frequency 

representations, although it has a higher computational cost than DWT.  At the same time, we 

observe that various modified versions of the Wavelet Transform have been used in extracting 

features from meditation EEG data. Some examples include Stationary Wavelet Transform 

[152], Flexible Analytic Wavelet Transform [184], Complex Morlet Wavelet Convolution [29], 

Morlet Wavelet Transform [164], Morlet Wavelet Decomposition [185], Complex Continuous 

Wavelet Coherence [183, 186], and Wavelet Coherence [187]. 

The Stockwell Transform [33, 91] is another method used to analyse EEG data in both time and 

frequency simultaneously to obtain TFD. It combines features from the Fourier Transform and 

the Wavelet Transform. The method uses a frequency-dependent Gaussian window that changes 

in size, allowing it to capture high- and low-frequency activities very efficiently. 

The Hilbert Transform [138, 188–190] closely aligns with time-frequency analysis techniques, 

as it produces an analytic signal for an EEG, from which amplitude and phase features are 

derived. The amplitude indicates how strong the signal is at a given time, while the phase shows 

the location in the signal cycle. Although not a TFD method, combining the Hilbert Transform 

with band-pass filters can isolate specific frequency bands, such as Alpha and Theta, which can 

be used in further studies. The Hilbert-Huang Transform [188] is a powerful signal processing 

method used to analyse nonlinear and non-stationary signals, such as EEG, consisting of two 

steps. In the first step, Empirical Mode Decomposition is performed, breaking the EEG signal 



33 
 

into simpler components called Intrinsic Mode Functions. In the second step, the Hilbert 

Transform is applied to these Intrinsic Mode Functions to obtain the amplitude feature set. 

The Spectral Entropy [191] is a feature extraction method applied to EEG data to quantify the 

complexity and unpredictability of the power distribution across different frequency bands. 

Spectral Entropy for an EEG dataset is calculated by first using Fourier or Wavelet Transform 

to calculate the PSD, and then taking the Shannon Entropy of the normalised PSD, where 

Spectral Entropy is a value ranging between 0 and 1. A high Spectral Entropy indicates that the 

EEG signal is complex, and the brain is under stress or working h