Inter-Subject Correlation Analysis in Music-Induced EEG Responses

An exploration of neural synchronization patterns in participants with varying levels of agreement on musical preferences, conducted as part of an independent study with the IIIT-H Music Cognition Group with the support and guidance of Prof. Vinoo Alluri.

Introduction

This independent study investigates how inter-subject correlation (ISC) in EEG responses varies between groups of participants who agree versus disagree on their subjective ratings of music. Using the DEAP dataset, we identified participant groups with high and low agreement on valence and liking ratings, then analyzed their neural synchronization patterns across different frequency bands and brain regions.

Part 1: Exploratory Data Analysis

Dataset Overview

The DEAP (Database for Emotion Analysis using Physiological signals) dataset provided the foundation for this study:

• 40 music videos (63 seconds each)
• 32 participants
• EEG recorded at 128 Hz → 8,064 timepoints per song
• 32 EEG channels categorized into 10 functional brain regions

Brain Region Classification

The 32 EEG channels were organized into the following regions:

region_dict = {
    "Prefrontal": [1, 17, 2, 18],        # Fp1, Fp2, AF3, AF4
    "Frontal": [3, 20, 4, 21, 19],       # F3, F4, F7, F8, Fz
    "Central": [6, 23, 5, 22, 7, 25, 24], # FC1, FC2, FC5, FC6, C3, C4, Cz
    "Temporal": [8, 26, 9, 27],          # T7, T8, CP5, CP6
    "Parietal": [11, 29, 12, 30, 16, 10, 28], # P3, P4, P7, P8, Pz, CP1, CP2
    "Occipital": [14, 32, 15, 13, 31],   # O1, O2, Oz, PO3, PO4
    "LeftHemisphere": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14],
    "RightHemisphere": [17, 18, 20, 21, 22, 23, 25, 26, 27, 28, 29, 30, 31, 32],
    "Midline": [15, 16, 19, 24],         # Oz, Pz, Fz, Cz
    "Global": list(range(1, 33))         # All electrodes
}

Subjective Ratings

Each participant rated all 40 songs on four dimensions (scale 1-9):

• Valence (negative to positive emotion)
• Arousal (calm to excited)
• Dominance (submissive to dominant)
• Liking (dislike to like)

For this study, we focused on Valence and Liking as the primary indicators of musical preference.

Correlation Analysis

Initial correlation matrices revealed participant-by-participant agreement patterns:

Key Finding: Participants 1 and 16 showed extremely low correlations with others and were excluded as outliers from subsequent analyses.

Computing Agreement and Disagreement Groups

Graph Theory Approach

Using graph theory and clique detection, we identified groups of participants with high mutual agreement and low mutual agreement on musical preferences.

Agreement Group Selection

Methodology:

1. Constructed a graph where:
   • Nodes = participants
   • Edges = participants with correlation ≥ 0.6 (valence) AND ≥ 0.5 (liking)
2. Identified cliques (fully connected subgraphs) with minimum 3 participants
3. Ranked cliques by scoring system:

Best Agreement Clique:

• Participants: 14, 15, 19, 20, 22, 32
• Average Score: 0.723
• Group Size: 6 participants

Disagreement Group Selection

An exhaustive search on the remaining 24 participants (excluding outliers 1, 16 and the 6 agreement participants) identified the group with the lowest mutual correlation scores.

Worst Disagreement Group:

• Participants: 6, 7, 9, 23, 27, 30
• Average Score: 0.193
• Group Size: 6 participants

Group Validation

The two groups showed clear separation in their correlation patterns:

Part 2: Inter-Subject Correlation (ISC) Computation

Methodology Overview

The ISC computation pipeline consisted of the following steps:

1. Load EEG data for specific song, group, frequency band, and region
2. Compute spectral power using Welch’s method
3. Create correlation matrices between participants
4. Calculate ISC statistics (min, max, mean, std)

Frequency Bands of Interest

freq_band_dict = {
    "Theta": (4, 8),      # Associated with memory and emotion
    "Alpha": (8, 13),     # Associated with attention and relaxation
    "Beta": (13, 30),     # Associated with active thinking
    "Gamma": (30, 45),    # Associated with cognitive processing
    "All": (4, 45)        # Broadband analysis
}

Data Preprocessing Challenges

Challenge 1: Electrode Labeling Inconsistency

Problem: Participants 1-22 used Twente channel ordering, while participants 23-32 used Geneva ordering.

Solution: Applied mapping table to convert all data to Geneva format for consistency.

Challenge 2: Trial Order Randomization

Problem: Each participant experienced the 40 songs in a different random order.

Solution: Used the Experiment_id column in participant_ratings.csv to correctly map trial numbers to song identities for each participant.

Spectral Power Estimation: Welch’s Method

What is Welch’s Method?

Welch’s method estimates the power spectral density (PSD) of a signal by:

1. Splitting the signal into overlapping windows
2. Applying a windowing function (Hamming) to each segment
3. Computing periodograms (FFT-based power estimates) for each window
4. Averaging periodograms to reduce noise

Parameters Used:

• Sampling rate: 128 Hz
• Signal duration: 63 seconds (8,064 timepoints)
• Window size: 2 seconds (256 samples)
• Hop length: 0.5 seconds (64 samples)
• Result: 123 windows per recording

ISC Calculation Process

For each combination of (song, region, frequency band, group):

1. Load and filter EEG:
   • Extract relevant channels for the brain region
   • Apply bandpass filter for frequency band
   • Stack across participants: (n_channels, n_timepoints, n_participants)
2. Compute spectral power:
   • Apply Welch’s method per channel and participant
   • Result: (n_channels, 123 windows, n_participants)
3. Create correlation matrices:
   • For each channel, compute pairwise Pearson correlations between all participant PSD time series
   • Result: (n_participants × n_participants) matrix per channel
4. Extract ISC values:
   • Take upper triangle of correlation matrix (excluding diagonal)
   • Compute statistics: min ISC, max ISC, mean ISC, std ISC

Computational Scope

Total combinations analyzed: 4,000

• 40 songs × 10 regions × 5 frequency bands × 2 groups = 4,000

Each combination produced one row in the output CSV:

Group, Song_Index, Frequency_Band, Region, Mean_ISC, Min_ISC, Max_ISC, Std_ISC

Part 3: Statistical Analysis of ISC

Analysis Framework

The mean ISC values from 4,000 combinations were aggregated for statistical testing:

• Statistical unit: mean ISC (origin-independent)
• Frequency bands: 5 (Theta, Alpha, Beta, Gamma, All)
• Brain regions: 10
• Analysis combinations: 50 (5 bands × 10 regions)
• Sample size per combination: 40 (one per song) for each group

Independent T-Tests

Hypothesis:

• Null (H₀): Agreement group ISC mean = Disagreement group ISC mean
• Alternative (H₁): The means of the two groups differ (two-tailed)

T-statistic formula:

Where:

• $\bar{X}_1, \bar{X}_2$ = sample means
• $s_1^2, s_2^2$ = sample variances
• $n_1, n_2$ = sample sizes (both = 40)

Degrees of freedom (Welch-Satterthwaite equation):

Results: 24 out of 50 combinations showed significant differences (p < 0.05)

Effect Size: Cohen’s d

Why effect size matters:

While p-values indicate whether an effect exists, Cohen’s d quantifies how large that effect is—crucial for interpreting practical significance in neuroscience.

Calculation:

For equal-sized groups, Cohen’s d can be approximated from the t-statistic:

Where t = t-statistic and n = 40 (observations per group)

Interpretation guidelines:

Key Findings from Effect Size Analysis

Top 5 Highest Effect Sizes:

Key Insights:

• Most differences showed small effects (d = 0.2-0.5), typical in naturalistic EEG with high variability
• Moderate effects (d ≥ 0.5) appeared exclusively in the Alpha band, particularly over frontal regions
• Alpha-band ISC appears most sensitive to shared cognitive states related to attention and engagement

Statistical Power Analysis

What is statistical power?

Power is the probability of detecting a true effect when it exists (correctly rejecting H₀). It depends on:

• Effect size (Cohen’s d)
• Sample size (n = 40 per group)
• Significance level (α = 0.05)

Power calculation:

where β = probability of Type II error

Computed using TTestIndPower from statsmodels based on Cohen’s d, sample size, and α = 0.05 (two-tailed).

Top 5 Highest Power Values:

Key Insights:

• Moderate power (0.3-0.6) for most tests—expected given small-to-medium effect sizes
• Highest power in Alpha band (>0.6 for Prefrontal, Frontal, Left Hemisphere)
• Lowest power in some Theta and All-band combinations (<0.3), indicating marginal detectability

Complete Results: 24 Significant Combinations

Theta Band (7 significant regions):

Alpha Band (9 significant regions):

All Band (8 significant regions):

Part 4: Complementary Per-Channel ISC Analysis

Motivation for Channel-Level Analysis

While region-based ISC (Part 2) provided broad insights into brain area differences, it masked finer spatial variations within regions. The per-channel approach offers:

• Higher spatial resolution: ISC computed for each of 32 electrodes independently
• Visual interpretability: Results visualized on topographic scalp maps
• Consistency verification: Identifies which specific electrodes drive regional effects

Computation Methodology

Input format: For each (Group, Song, Frequency Band) combination:

• EEG shape: (32 channels, 8064 timepoints, 6 participants)
• All 32 electrodes loaded per participant (no regional grouping)
• Same spectral power and ISC computation as Part 2, but per individual channel

Output structure:

Group, Song_Index, Frequency_Band, ISC1, ISC2, ..., ISC32

• One row per combination
• Total rows: 2 groups × 40 songs × 5 bands = 400
• Each ISC_i = mean ISC for channel i across all participant pairs

Topographic Visualization

To interpret spatial patterns of ISC differences:

1. Average ISC per channel across all 40 songs for each group and frequency band
2. Plot on scalp using MNE’s topomap visualization (biosemi32 montage)
3. Compare Agreement vs. Disagreement groups side-by-side

Benefits:

• Provides spatial context to statistical results from Part 3
• If “Parietal-Alpha” was significant in region-based analysis, topomaps show which parietal electrodes drove the effect
• Reveals gradients and hotspots of neural synchronization

Part 5: Comprehensive Visualizations

Heatmap: T-Test Results Across All Combinations

This heatmap displays the difference in mean ISC between Agreement and Disagreement groups for all 50 (Region × Band) combinations:

Legend:

• Color intensity = magnitude of ISC difference (Agreement - Disagreement)
• * denotes statistically significant p-value (p < 0.05)
• Difference in means is directly proportional to Cohen’s d (effect size)

Key Observations:

• Alpha band shows the most widespread and strongest differences
• Prefrontal, Frontal, and Left Hemisphere regions exhibit largest effects
• Beta and Gamma bands show minimal differences

Topographic Maps: Alpha Band Comparison

The Alpha band was the only frequency range with consistently significant combinations showing moderate effect sizes and statistical power. Below are comparative topomaps for Agreement vs. Disagreement groups:

Visual Insights:

• Agreement group shows higher ISC across frontal and parietal regions
• Particularly pronounced in prefrontal (Fp1, Fp2, AF3, AF4) and frontal (F3, F4, Fz) electrodes
• Disagreement group shows more uniform, lower ISC across the scalp
• Spatial pattern confirms region-based statistical findings

Final Interpretations

Statistical Summary

1. Alpha Band Dominance

Finding: The Alpha band (8-13 Hz) showed the strongest and most consistent differences between agreement and disagreement groups across nearly all brain regions.

Evidence:

• Highly significant p-values across all regions (all p < 0.05, most p < 0.01)
• Largest effect sizes of any frequency band
• Highest statistical power (up to 0.68 for Prefrontal)

2. Small but Meaningful Effects

Finding: Most group differences showed small effects (Cohen’s d = 0.2-0.5), which is expected and appropriate in naturalistic EEG settings.

Interpretation:

• High inter-individual variability is inherent in real-world neural data
• Small effects can still be meaningful when consistently observed across multiple conditions
• Moderate effects in Alpha band suggest this frequency range is particularly sensitive to shared cognitive states

3. Spatial Patterns

Finding: Moderate effects were concentrated in frontal and parietal regions, with strong left hemisphere involvement.

Topographic confirmation:

• Agreement group participants showed elevated ISC particularly in prefrontal (Fp1, Fp2, AF3, AF4) and frontal (F3, F4, Fz) electrodes
• Spatial gradient visible from front to back, with strongest effects anteriorly
• Consistent with region-based statistical tests

Cognitive Interpretations

While formal cognitive conclusions require further validation, these patterns suggest several plausible interpretations:

Alpha Oscillations and Attention

Alpha-band activity (8-13 Hz) is strongly associated with:

• Attention allocation and selective inhibition
• Working memory maintenance
• Top-down cognitive control

Implication: Participants who agree on musical preferences may exhibit more similar patterns of attention deployment and cognitive engagement during listening.

Frontal Involvement

Prefrontal and frontal regions are critical for:

• Aesthetic judgment and preference formation
• Emotional evaluation
• Executive control during music listening

Implication: Neural synchronization in these regions may reflect shared evaluative processes and emotional responses to music.

Left Hemisphere Lateralization

Left-hemisphere dominance in the observed effects may relate to:

• Temporal processing and rhythm perception
• Language-like analytical processing of music
• Positive emotional valence (left-frontal approach motivation)

Implication: Agreement on musical preferences may involve similar left-lateralized cognitive strategies for music processing.

Methodological Strengths

1. Robust group identification: Graph-theoretic clique detection ensured maximal within-group agreement and between-group separation
2. Comprehensive frequency analysis: Five frequency bands provided complete spectral coverage
3. Multiple spatial scales: Both region-based and per-channel analyses offered complementary insights
4. Rigorous statistics: Effect sizes and power analyses contextualized significance tests
5. Visual validation: Topographic maps provided intuitive spatial confirmation of statistical results

Limitations and Future Directions

Current limitations:

• Sample size (n=6 per group) limited by strict agreement/disagreement criteria
• Cross-sectional design prevents causal inference
• Musical stimuli varied in genre, tempo, and structure

Future research could:

• Expand to larger participant pools with varied musical backgrounds
• Control for specific musical features (tempo, key, complexity)
• Investigate other frequency bands (Beta, Gamma) with targeted hypotheses
• Explore directed connectivity measures (e.g., Granger causality)
• Incorporate behavioral measures of engagement and attention

Conclusion

This study demonstrated that inter-subject correlation in Alpha-band EEG activity differs significantly between groups with high versus low agreement on musical preferences, particularly in prefrontal, frontal, and parietal regions. These findings suggest that shared musical preferences may be associated with similar patterns of neural synchronization during listening, potentially reflecting common cognitive and emotional processing strategies.

The Alpha band’s unique sensitivity to these group differences highlights the importance of attentional and evaluative processes in shaping individual music preferences. Future work combining neural, behavioral, and computational approaches will be essential for fully understanding the complex relationship between brain activity and musical experience.

Acknowledgments

This work was conducted as part of an independent study with the IIIT-Hyderabad Music Cognition Group under the guidance of Prof. Vinoo Alluri. I am deeply grateful to Prof. Alluri for her invaluable support, insights, and mentorship throughout this project, and to the members of the Music Cognition Group for their feedback and encouragement.

Technical Implementation

Tools and Libraries:

• Python 3.x for all analyses
• NumPy & SciPy for signal processing
• Pandas for data manipulation
• MNE-Python for EEG visualization
• Statsmodels for statistical testing
• Matplotlib & Seaborn for visualization
• NetworkX for graph-theoretic clique detection

Dataset:

• DEAP (Database for Emotion Analysis using Physiological signals)
• Available at: http://www.eecs.qmul.ac.uk/mmv/datasets/deap/

For questions or collaboration opportunities, feel free to reach out.