Skip to content
Music×EEG

B.Tech Research Synthesis · 2 Independent EEG Datasets

Can Music Improve Cognitive Working Capability?

Two EEG datasets. One honest answer: not yet established.

We investigated two complementary datasets, analysed and trained completely separately. A 50-person single-channel NeuroSky MindWave arithmetic cohort tests what one forehead electrode can and cannot decode — it has no music at all. A 10-person, 4-channel PhysioNet N-back cohort adds four music conditions, but in a fixed order that confounds music with practice.

Research scope · dual-dataset architecture

Separate pipelines · no pooled training
Dataset 1N = 50

NeuroSky MindWave

  • · Single dry electrode, FP1 · TGAM1 ASIC · 512 Hz · raw ADC counts
  • · Unpaced mental arithmetic: Rest (eyes closed) vs Medium vs Hard (eyes open)
  • · Zero music conditions — a pure workload baseline
  • · Role: legacy audit, leak-free baseline, alpha ablation
Open MindWave →
Dataset 2N = 10

PhysioNet Experiment 1

  • · 4 dry electrodes (TP9, AF7, AF8, TP10) · Muse · 256 Hz · calibrated µV
  • · Visual N-back: 1-back (low load) vs 3-back (high load), 64 blocks per person
  • · 4 music conditions in a fixed order: No Music → Relaxing → Stressing → AI
  • · Role: synchronisation, personalisation, confound audit
Open PhysioNet →

Consolidated synthesis (conclusions only — not data)

  • → Generic cross-subject models sit near chance on fine-grained load
  • → Personal baseline calibration gives significant gains
  • → Bandpower ratios beat raw power; single-trial ERPs fail
  • → Uncounterbalanced music designs are confounded by practice
clean 1-s MindWave chunks MW
9,074
stimulus-locked PhysioNet epochs PN
14,080
balanced-accuracy inflation from chunk-level leakage MW
+11.33 pp
personalized gain over generic models PN
+4.45 to +5.10 pp

Source: reports/CONSOLIDATED_RESEARCH_FINDINGS.md · §A–§C

How the pipeline works

Three rules, no shortcuts

Both datasets run through the same discipline — different code paths, same guarantees.

Step 1

Zero-Phase DSP

1–40 Hz Butterworth · SOS · sosfiltfilt

Order-4 Butterworth bandpass in second-order sections, run forward-backward so phase is identically zero. MindWave filters inside continuous segments; PhysioNet filters the full ~87-min recording on two tracks (1–40 Hz bandpower, 0.5–30 Hz ERP) before epoching.

Step 2

Leak-Free Splits & Calibration

GroupKFold · LOSO · chronological blocks

Generic models never see a test subject (GroupKFold / LOSO by participant). Personalized models calibrate only on early blocks or chunks and are evaluated strictly on later ones. Scalers are fit on training or calibration data only.

Step 3

Honest Neurophysiological Benchmarks

Effect sizes · permutation · Holm / BH

Every claim carries d_z, non-parametric tests and multiple-comparison control — and is reported alongside the negative controls (Medium vs Hard, ERP subsystem, music conditions) that bound it.

Core findings across both cohorts

What the evidence actually supports

Four observations that hold in both datasets — each paired with the method that produced it.

01

≈ 51%

generic 1-back vs 3-back

Inter-Subject Variance Barrier

Generic, uncalibrated models fail on fine-grained load. PhysioNet 1-back vs 3-back sits at 50.97–51.19% balanced accuracy across RF, LogReg and SVM; MindWave Medium vs Hard Math sits at ~52–53%, below the 56.99% majority baseline.

GroupKFoldLOSOBoth datasets
02

+4.45 pp

RF personalized gain, p = 0.0025

The Personalization Imperative

Calibrating against each person’s own baseline wins in both cohorts: +3.75 pp from MindWave resting calibration; +4.45 pp (RF, d_z = 1.31, Holm p_adj = 0.0075) to +5.10 pp (SVM) from PhysioNet 20% chronological calibration. 9 of 10 PhysioNet participants improved.

Chronological splitHolm–BonferroniBH-FDR
03

56.25% vs 50.26%

band_ratios vs erp

Spectral Ratios Beat Raw Power & Single-Trial ERPs

theta/alpha and theta/beta style ratios self-normalise across electrode impedance. PhysioNet band_ratios rank #1 (56.25%) while the ERP subsystem fails at chance (50.26%); in MindWave, theta_alpha_z is the top feature of the full model.

Feature ablationθ/αθ/β
04

p = 0.207

music-condition RM-ANOVA (synthesis)

Experimental Confounds Audit

MindWave Rest is eyes-closed and Math eyes-open (Berger effect) — alpha drops at eye opening (d_z = −1.378) but is flat from Medium to Hard (p = 0.2553). PhysioNet runs No Music → Relaxing → Stressing → AI Music in a fixed order, so accuracy 36.51% → 51.79% tracks practice.

Negative controlFixed orderBerger effect

Reading the accuracies correctly

MindWave's ~72–73% separates eyes-closed rest from active arithmetic. PhysioNet's ~55.6–56.1% separates two eyes-open active workload levels after personal calibration. Different contrasts, different difficulty — the numbers are not a ranking of headsets or models.

Claims boundary · read before citing

What we have NOT established

These seven boundaries are a result, not a footnote. Most importantly: we have not shown that music causally improves cognitive working capability.

  1. ✕ 01

    We have not established that music causally improves cognitive working capability.

    MindWave has no music conditions. In PhysioNet every participant heard the conditions in the same order; the 36.5% → 51.8% gain is a learning curve, the music-condition RM-ANOVA is non-significant (p = 0.207) and no EEG band differs across conditions after FDR (all q ≥ 0.90).

  2. ✕ 02

    We have not established that EEG-guided music selection reduces perceived mental fatigue.

    Neither dataset recorded NASA-TLX, the Karolinska Sleepiness Scale or any subjective fatigue scale across blocks.

  3. ✕ 03

    We have not established that our models generalize to unseen users.

    Generic models on unseen participants scored 50.0%–51.2%. The validated models are personalized and need calibration data from each user.

  4. ✕ 04

    We have not established that 4-channel Muse findings transfer to single-channel MindWave hardware.

    PhysioNet’s strongest effects were temporoparietal (TP10, d_z = 1.26); MindWave records only FP1, in unscaled counts, without mastoid references.

  5. ✕ 05

    We have not established that offline classifications translate to real-time, closed-loop operation.

    All evaluations used pre-recorded, zero-phase filtered epochs. Streaming needs causal filtering and faces buffer latency, motion and impedance drift.

  6. ✕ 06

    We have not established an optimal calibration duration.

    The 20% budget (12.4 min, 55.64%) and the 50% budget (57.06% on the common set) describe an observed trend on retrospective data, not a proven optimum.

  7. ✕ 07

    We have not established that heuristic artifact screening removes all physiological contaminants.

    Neither dataset has EOG/EMG channels; blinks, saccades and frontalis/temporalis tone can survive amplitude thresholds such as 200 µV.

Source: reports/CONSOLIDATED_RESEARCH_FINDINGS.md · Section F

How Semester 8 tests this →

Fair scientific questions

Questions a reviewer would ask