Skip to content
Music×EEG
Dataset 2N = 10 · TP9 · AF7 · AF8 · TP10 · 256 Hz · µVDOI 10.13026/jr6f-wc85

PhysioNet Experiment 1: N-back under music, in a fixed order

Ten participants each completed 64 visual N-back blocks across four continuous auditory sessions while wearing a 4-channel Muse headband. It is our primary benchmark for synchronisation, psychometrics and personalized decoding — and the clearest demonstration of why an uncounterbalanced design cannot answer the music question.

14,080
stimulus-locked 2-s epochs
0.98 ms
mean alignment error (max 2.0 ms)
97.81%
eligible after screening (13,772)
+4.45 pp
RF personalized gain (p = 0.0025)

Phases 2–3B · Experimental structure

4 sessions × 16 blocks × 22 trials = 1,408 per person

Every participant heard the conditions in the same order. Practice blocks (1–2) and the mid-session rest blocks (11, 28, 45, 62) are excluded, which is why block IDs run 3–70.

  1. S1

    No Music

    Cognitive baseline

    blocks 3–10, 12–19

  2. S2

    Relaxing

    Participant-selected calming music

    blocks 20–27, 29–36

  3. S3

    Stressing

    Participant-selected upbeat / exciting music

    blocks 37–44, 46–53

  4. S4

    AI Music

    Newly synthesised (AI-generated) relaxing music

    blocks 54–61, 63–70

Inter-session rests: 3, 6 and 3 minutes. Each session: 8 × 1-back and 8 × 3-back blocks. Protocol ≈ 85 min; mean continuous EEG 87.1 min.

Synchronisation (Phase 3A)

Sub-sample stimulus alignment

  • · Mean error 0.98 ± 0.7 ms, max 2.000 ms
  • · 14,080 / 14,080 experimental trials mapped (100%)
  • · Zero transmission gaps > 4.0 ms anywhere in the cohort
  • · Timestamps quantised to alternating 3.0 / 4.0 ms (256 Hz raster)
  • · A3, A5 and A10 repeat a voltage sample every 1,285 samples (≈ 5.02 s). Raw data are left unmodified; an N = 7 sensitivity cohort excludes them.

Source: PROJECT_RESEARCH_LOG.md · §8.1–8.2

Preprocessing (Phase 3B)

Two filter tracks, one fallback rule

  • · Track A — Cognitive bandpower: 1.0–40.0 Hz — Rhythm power, entropy, workload ratios
  • · Track B — ERP / time-domain: 0.5–30.0 Hz — Evoked deflections, temporal morphology
  • · Butterworth prototype N = 4 in SOS form, zero-phase sosfiltfilt on the continuous ~87-min recording before epoching (≈ 48 dB/oct single-pass, ≈ 96 dB/oct forward-backward per edge).
  • · A4: AF7 saturated → AF8 single-channel fallback
  • · A9: AF8 (and TP10) saturated → AF7 single-channel fallback
  • · frontal_alpha_asymmetry computed only when both frontal electrodes are usable; NaN for A4, A9 and artifact trials.

Source: reports/physionet/experiment_1/eeg_preprocessing_and_feature_extraction.md

Five-tier channel-quality taxonomy
TierRuleMeaning
1 · SATURATION_CLIPPING≤ −999 µV or ≥ +999 µVFatal — ADC rail limit
2 · FLATLINE_DISCONNECTEDσ < 0.5 µVFatal — detached electrode
3 · GROSS_MOTION_SLIPPTP > 300 µVGross movement
4 · SUSPECTED_OCULAR_EMG100 µV < PTP ≤ 300 µVBlinks / temporalis EMG
5 · PHYSIOLOGICAL_RANGE_CANDIDATEPTP ≤ 100 µVHeuristic candidate — artifacts can persist
Model eligibility screen (Frontal PTP ≤ 200 µV on AF7/AF8 (model eligibility screen)) — 308 excluded (2.19%), 13,772 eligible
IDExcludedEligibleCalib. (20%)Eval.Calib. min
A160 (4.26%)1,3482791,06912.4
A235 (2.49%)1,3732831,09012.3
A30 (0.00%)1,4082861,12212.4
A490 (6.39%)1,3182451,07312.5
A58 (0.57%)1,4002861,11412.3
A64 (0.28%)1,4042841,12012.4
A744 (3.12%)1,3642701,09412.6
A824 (1.70%)1,3842841,10012.3
A940 (2.84%)1,3682761,09212.1
A103 (0.21%)1,4052861,11912.4

Highlighted rows: participants relying on single-channel frontal fallback. The 200 µV screen is a project heuristic; ocular and EMG artifacts can persist below it.

Source: reports/physionet/experiment_1/phase4_final_validation.md · §2.1

Phase 2 · Behavioural psychometrics

The load works. The music question can't be answered.

Working-memory load effect · valid

d′ 1.817 → 1.11

1-back d′1.817 ± 0.882
3-back d′1.110 ± 0.684

F(1, 9) = 17.14, p = 0.0025, η²p = 0.656; Wilcoxon W = 3.0, p = 0.0098. Accuracy 48.89% → 38.38% (p = 0.0488 (Wilcoxon W = 8.0)). Correct-trial RT 395.8 → 389 ms (n.s., p = 0.117). The 1-back / 3-back labels are a genuine, verified workload contrast. (d′ bars drawn on a 0–3 scale.)

Music conditions · confounded

Accuracy rises in session order

  1. S1 · No Music36.51%
  2. S2 · Relaxing40.99%
  3. S3 · Stressing45.26%
  4. S4 · AI Music51.79%

Every participant ran No Music → Relaxing → Stressing → AI Music. Condition is perfectly collinear with elapsed time and practice over ~85 minutes, so this monotonic rise is a practice curve, not evidence that music helps. SD across participants is ~30 pp in every session.

Planned pairwise music contrasts (paired t, Holm-corrected) — none survive
ContrastΔ acc (pp)p rawp Holmdz
Relaxing vs. No Music+4.490.20910.50600.428
Stressing vs. No Music+8.750.16870.50600.473
AI Music vs. No Music+15.280.03300.17720.796
Stressing vs. Relaxing+4.260.31080.50600.340
AI Music vs. Relaxing+10.800.02950.17720.817
AI Music vs. Stressing+6.530.07370.29490.640

Source reconciliation: two RM-ANOVA results exist in the repository

  • Synthesis (consolidated report & research log): RM-ANOVA across music conditions F(3, 27) = 1.62, p = 0.207.
  • Primary behavioural report: 2 (load) × 4 (condition) RM-ANOVA on accuracy, condition main effect F(3, 27) = 4.04, p = 0.0171, η²p = 0.31.
  • Both agree: Friedman χ²(3) = 5.67, p = 0.129; no pairwise contrast survives Holm (padj ≥ 0.177); EEG shows no condition difference after FDR (all q ≥ 0.9). Even a significant condition effect would be uninterpretable under a fixed order.

A4

75.9% omissions

mean accuracy 18.25% — disengagement / task difficulty; RT metrics handled cautiously

A6

96.2% omissions

mean accuracy 3.85% — disengagement / task difficulty; RT metrics handled cautiously

A10

95.7% omissions

mean accuracy 5.11% — disengagement / task difficulty; RT metrics handled cautiously

Source: reports/physionet/experiment_1/behavioural_analysis.md · §3–5

Phase 3C · Exploratory EEG neuromarkers

Large effects, zero FDR survivors

Paired participant-level 1-back vs 3-back comparisons. Effect sizes are large, but with N = 10 and 42 features tested, none survive Benjamini–Hochberg correction.

Theta modulation by working-memory load (exploratory)
FeatureSiteChangedzp (uncorr.)FDR q
TP10_rel_thetaRight temporoparietal+6.87%+1.260.00390.1641
TP9_rel_thetaLeft temporoparietal+6.92%+1.040.00980.2051
AF7_rel_thetaLeft anterior-frontal+9.28%+0.680.04880.5496
AF8_theta_alphaRight anterior-frontal+7.40%+0.690.05670.5496

FDR reality check: q ≥ 0.164 for every feature

These are promising candidate markers, not confirmed neurophysiology. Temporoparietal sites (TP9/TP10) show larger effects than frontal sites.

Trait, not state

ρ = -0.8788

Mean baseline frontal theta vs d′ across participants (Spearman, N = 10, p = 0.0008; N = 8 subset ρ = -0.9286, p = 0.0009). Leave-one-participant-out ρ ∈ [−0.9167, −0.8333] (all p ≤ 0.0068). Within-subject session fluctuations do not track performance (|r_rm| < 0.20).

Robustness

  • · N = 10 vs N = 7: Excluding A3, A5, A10 preserves 100% of effect directions; |Δd_z| < 0.18.
  • · Relaxing 200 µV → 300 µV adds only +124 trials (+0.88%); effect sizes shift by < 0.01 d_z.

Source: reports/physionet/experiment_1/key_findings.md

Phase 4 · Personalized cognitive-load classification

Generic fails. Personalized wins — modestly.

1-back vs 3-back with perfectly balanced classes (chance = 50%). Personalized models calibrate on each person's first 13 blocks (20%, 2,779 trials, ≈ 12.4 min) and are tested on their remaining 51 blocks (10,993 trials) — the same blocks used to score the generic models.

Generic vs personalized (± SD)

Generic vs personalized balanced accuracy
ConditionValue (%)Spread
Logistic Regression · generic50.98± 3.33 SD
Linear SVM · generic50.97± 3.52 SD
Random Forest · generic51.19± 2.55 SD
Random Forest · personalized55.64± 4.42 SD
Logistic Regression · personalized55.96± 4.74 SD
Linear SVM · personalized56.07± 4.71 SD
50% chanceerror bar = reported SD or 95% CI

Paired significance (df = 9)

ModelΔ pp95% CIdzpHolmBH q
Random Forest+4.45[2.02, 6.89]1.310.00250.00750.0075
Logistic Regression+4.98[1.29, 8.68]0.960.01380.02280.0138
Linear SVM+5.10[1.46, 8.75]1.000.01140.02280.0138

Wilcoxon signed-rank agrees: RF W = 2.0, p = 0.0059 (Holm 0.0176); LogReg and SVM W = 3.0, p = 0.0098 (Holm 0.0195); BH q = 0.0098 for all three. All three architectures survive Holm–Bonferroni and Benjamini–Hochberg control.

Condition-matched generic baseline

Generic RF trained only on Session 1 (No Music): 50.00% ± 3.82. Personalized still wins by +5.63 ± 4.64 pp (dz = 1.21, t(9) = 3.84, p = 0.0040; Wilcoxon W = 0.0, p = 0.0020). Generic training restricted to Session 1 (No Music) removes one distribution mismatch, but evaluation spans Sessions 2–4, so order, practice and fatigue remain potential confounders.

Per-participant gain · Random Forest

10 of 10 improved over generic-all; A2 barely moved (+0.38 pp).

IDGenericCond.-matchedPersonalizedΔ vs genericHeld-out N
A148.27%49.77%55.77%
+7.50
1,069
A250.32%49.86%50.70%
+0.38
1,090
A354.12%47.38%60.77%
+6.65
1,122
A448.65%47.39%56.84%
+8.19
1,073
A552.79%51.04%57.24%
+4.45
1,114
A650.31%49.11%50.93%
+0.62
1,120
A754.89%56.90%64.88%
+9.99
1,094
A854.40%56.59%55.23%
+0.83
1,100
A949.50%43.92%50.87%
+1.37
1,092
A1048.66%48.04%53.18%
+4.52
1,119

Source: reports/physionet/experiment_1/phase4_final_validation.md · §3.1

Calibration budget on a fixed test set

Sessions 3–4 (blocks 37–70), 6,888 trials (3,446 3-back / 3,442 1-back), RF ± SD.

Balanced accuracy by calibration budget on the fixed common evaluation set
ConditionValue (%)Spread
10% · 130 trials · 5.8 min52.19± 4.87 SD
20% · 278 trials · 12.4 min54.68± 4.62 SD
30% · 414 trials · 18.2 min55.77± 6.78 SD
50% · 689 trials · 31 min57.06± 5.62 SD
50% chanceerror bar = reported SD or 95% CI

Original-partition accuracies (test set shrinks as budget grows): 10% 53.64% · 20% 55.64% · 30% 56.68% · 50% 56.17%. An observed trend — not an optimal calibration duration.

Feature-group ablation (personalized 20%)

Balanced accuracy by feature group
ConditionValue (%)Spread
band_ratios (12)56.25± 5.30 SD
absolute_power (20)55.86± 5.46 SD
relative_power (20)55.82± 6.08 SD
time_domain (24)55.73± 5.06 SD
core_ml (27)55.66± 5.06 SD
temporoparietal (20)55.33± 5.16 SD
compact_rhythms (13)55.21± 4.82 SD
frontal_subsystem (21)53.67± 4.85 SD
erp (16)50.26± 2.91 SD
50% chanceerror bar = reported SD or 95% CI

12 self-normalising ratios beat the full set; single-trial ERP features are at chance. Normalisation fit on calibration blocks: Raw 53.49% · z-Score 55.66% · MinMax 55.98%.

What ≈ 55.6–56.1% means — and doesn't

A statistically robust but modest offline single-trial baseline. It is not proven workload detection for unseen users, not evidence of music-induced improvement, and not a real-time closed-loop BCI. Calibration blocks all fall in Session 1 (No Music), so temporal confounds remain.

Source: reports/physionet/experiment_1/phase4_modeling_results.md

Curated figures

PhysioNet diagnostics

Click any figure for the full-resolution viewer (← / → to step, Esc to close). All 42 PhysioNet figures →

PhysioNetSignal Processing & Quality

Participant recording timelines

Continuous EEG coverage across the ~85-minute protocol for all 10 participants: session blocks, 3- and 6-minute rest breaks, manual tags and the reconstructed A8 Session 2 onset.

PhysioNetSignal Processing & Quality

Dual-track SOS filter responses

Magnitude and phase for both tracks, with single-pass (~48 dB/oct) and forward-backward (~96 dB/oct) roll-off per transition edge.

PhysioNetSignal Processing & Quality

Cohort channel usability (97.8% frontal yield)

Retention per scalp channel and overall frontal epoch yield. Mastoid channels suffer temporalis tension; frontal fallback restores 97.8% (13,772 / 14,080) usable trials.

PhysioNetNeurophysiology & Behavior

Accuracy by music condition (fixed order)

Cohort accuracy across the four sessions: 36.51% → 40.99% → 45.26% → 51.79%. Because every participant ran the same order, this rise is collinear with practice and cannot be attributed to music.

PhysioNetNeurophysiology & Behavior

1-back vs 3-back behavioural contrast

Participant-level paired lines: d′ drops from 1.817 to 1.110 (F(1,9) = 17.14, p = 0.0025, η²p = 0.656) and accuracy from 48.89% to 38.38% under 3-back load.

PhysioNetNeurophysiology & Behavior

Practice curve vs flat EEG across sessions

Behavioural accuracy rises across the four fixed-order sessions while EEG rhythms stay flat (no condition difference survives FDR, all q ≥ 0.90) — the signature of a practice effect, not a music effect.

PhysioNetNeurophysiology & Behavior

Paired theta rise, 1-back → 3-back

Participant-level paired slopes for EEG features. TP10_rel_theta d_z = +1.26 (p = 0.0039 uncorrected). Exploratory: zero of 42 features survive BH-FDR (q ≥ 0.164).

PhysioNetNeurophysiology & Behavior

Frontal theta vs working-memory sensitivity

Between-subject scatter: mean frontal theta vs d′, Spearman ρ = −0.8788 (p = 0.0008, N = 10). Trait-level association only — within-subject fluctuations do not track performance (|r_rm| < 0.20).

PhysioNetMachine Learning & Ablation

Generic vs personalized accuracy

Generic All (51.2%), Generic Condition-Matched (50.0%) and Personalized (55.6%) RF balanced accuracy on identical held-out blocks (p = 0.0025, d_z = 1.31, Holm p_adj = 0.0075).

PhysioNetMachine Learning & Ablation

Per-participant calibration delta

Waterfall of ΔBalAcc = personalized − generic per participant (largest: A7 +9.99, A4 +8.19, A1 +7.50, A3 +6.65 pp; smallest: A2 +0.38 pp).

PhysioNetMachine Learning & Ablation

Calibration budget scaling

Accuracy across 10/20/30/50% budgets under original partitions and on the fixed common evaluation set (Sessions 3–4): 52.19% → 54.68% → 55.77% → 57.06%. An observed trend, not a proven optimum.

PhysioNetMachine Learning & Ablation

Feature-group ablation

Ranking of feature subsystems under personalized calibration: band_ratios first (56.25%), erp last at chance (50.26%).