Skip to content
Music×EEG

Methodology · reproducibility · next phase

Semester 8: a study that can answer the music question

Both Semester 7 datasets failed the causal test for design reasons — no music in one, a fixed order in the other. The prospective wearable study below is built to remove exactly those confounds.

Experimental blueprint

Four layers, every confound addressed

01

Hardware & streaming

  • Wearable headband (Muse 2 / Muse S, 4 channels: TP9, AF7, AF8, TP10)
  • Lab Streaming Layer (LSL) / Bluetooth into a Python real-time engine
  • Causal SOS IIR filtering and online 2-second buffer windowing
02

Design & controls

  • Visual N-back (1-back vs 3-back) with millisecond stimulus logging
  • Latin-square counterbalancing: Silence vs Subject-preferred relaxing vs Standard classical vs AI audio
  • Washout rest intervals (3–5 min) between conditions
  • Calibrated sound level (e.g. 65 dBA), identical over-ear headphones, quiet room
03

Resting-state calibration

  • 2 min eyes-closed + 2 min eyes-open rest BEFORE any cognitive task
  • Fit per-person MinMax / z-score baselines from rest (no task trials consumed)
  • Self-normalising ratios: theta/beta, theta/alpha, alpha/beta
04

Multimodal outcomes

  • Behaviour: accuracy, RT, RT variability, d′, criterion c, omission rate
  • Subjective: NASA-TLX after each condition; Karolinska Sleepiness Scale or VAS fatigue; liking / familiarity / distraction
  • EEG: frontal & temporoparietal theta/alpha, frontal alpha asymmetry, θ/β ratios

Why counterbalancing fixes the PhysioNet confound

Semester 7 · PhysioNet (fixed order)

Every participant, same sequence — music condition = time on task.

S1S2S3S4
P1NoneRelaxStressAI
P2NoneRelaxStressAI
P3NoneRelaxStressAI
P4NoneRelaxStressAI

Semester 8 · Latin square (illustrative 4 × 4)

Each condition appears once in every serial position, decoupling music from practice and fatigue.

S1S2S3S4
P1SilencePref.ClassicalAI
P2Pref.ClassicalAISilence
P3ClassicalAISilencePref.
P4AISilencePref.Classical

Pref. = subject-preferred relaxing; Classical = standard classical; AI = AI-generated audio. Plus 3–5 min silent washouts between conditions.

Retain from Semester 7

  • Personal baseline normalisation fit strictly on baseline data
  • Compact spectral ratios; drop single-trial ERP features from the online feature space
  • Chronological, leak-free evaluation (test strictly after calibration)
  • Wilcoxon + d_z + Holm–Bonferroni and Benjamini–Hochberg on every paired comparison

Must be validated on real hardware

  • Bluetooth/LSL streaming without dropouts or buffer stalls
  • Electrode contact and sweat-induced impedance drift over 45–60 min
  • Whether 2 min of rest calibrates as well as 12+ min of task blocks

Source: reports/CONSOLIDATED_RESEARCH_FINDINGS.md · Section G

Pre-stated hypotheses

H1–H4: open questions, not results

  1. H1

    Neurophysiological modulation

    Does personalized or relaxing background music attenuate the rise in frontal and temporoparietal theta/beta ratios during high-load N-back, compared with silence, under counterbalancing?

  2. H2

    Behavioural performance

    Does background music measurably improve d′ or reduce RT variability during sustained tasks once practice effects are controlled?

  3. H3

    Subjective fatigue

    Does auditory intervention reduce NASA-TLX workload and KSS fatigue over prolonged blocks compared with silence?

  4. H4

    Calibration efficacy

    Can a 2-minute pre-task resting recording match the single-trial classification performance of active task-based calibration (≈ 12.4 min)?

Reproducibility

Rebuild every number yourself

Global random seed 0 across models, splitters, bootstraps and permutations. Raw data are git-ignored and verified by checksum.

reproduction commands (repo root)bash
  1. $ python -m pytest tests/

    # Full unit-test suite (DSP, validation, PhysioNet, Phase 4 modeling)

  2. $ python src/03_verify_baseline.py data/processed/features.csv

    # Historical MindWave baseline — 13 regression checks

  3. $ python -m src.physionet.experiment1.behavioural_analysis

    # Behavioural pipeline & statistics

  4. $ python -m src.physionet.experiment1.generate_sync_figures

    # Phase 3A synchronisation figures (5)

  5. $ python -m src.physionet.experiment1.generate_preprocessing_figures

    # Phase 3B preprocessing figures (10)

  6. $ python -m src.physionet.experiment1.generate_phase3c_figures

    # Phase 3C statistics & ML figures (12)

  7. $ python -m src.physionet.experiment1.modeling.model_comparison

    # Phase 4 matched model evaluation & statistics

  8. $ python -m src.physionet.experiment1.modeling.generate_phase4_figures

    # Phase 4 figures (9)

Test-suite status

Documented (with all data present)
65 / 65 pytest passing; MindWave baseline verifier 13 / 13 — per PROJECT_RESEARCH_LOG.md §18 and phase4_final_validation.md §6.
MindWave suite
18 / 18 passing (test_dsp.py, test_validation.py, test_phase9.py) — re-run while building this site.
PhysioNet suite on a fresh clone
Data-dependent tests need the git-ignored PhysioNet raw files (706,091,119 bytes) and processed tables in data/; without them they fail at load time.

SHA-256 provenance

MindWave raw CSV
merged_eeg_datas_50_subjects.csv · 19,029,871 bytes
5d890ac079aa3512f4ba338accee00d1ed960d2db2c8f0261baa363b1fffce5b
MindWave features.csv
4,703,782 bytes
ea41de332f7d5a3054b00e421fcfbdabfd50562e14880bc153084d9387b6f939
PhysioNet Experiment 1
30 / 30 raw files match the official SHA256SUMS.txt · 706,091,119 bytes

Repository index

Synthesis

MindWave

Source: PROJECT_RESEARCH_LOG.md · §15, §18