中文版 | English

MODMA EEG Depression Classification

MDD vs HC binary classification using 128-channel ERP data from the MODMA dataset. Statistically significant results achieved through QC repair, P300 feature extraction, electrode selection, time-window ablation, and feature importance analysis.

Key Findings

Model	Dims	BA	p-value	Significant
hcue P300 mean amplitude	128	0.670	0.022	Yes
Parietal subset (Pz/P3/P4/Cz/CPz)	5	0.634	0.045	Yes
250-300ms time window	128	0.673	0.018	Yes

• hcue dot-probe P300 mean amplitude is the only statistically significant MDD classification feature
• E55 (Cz-adjacent) has the highest channel weight, with positive deflection toward HC
• Feature enrichment and multi-condition fusion both degraded performance
• Cross-dataset generalization (TDBRAIN resting-state) failed (BA=0.500)

Dataset

MODMA 128-channel ERP dataset:

• 53 subjects (24 MDD + 29 HC), 128-ch EGI HydroCel, 250Hz
• Dot-probe emotional attention bias paradigm (hcue/fcue/scue conditions)
• 52 subjects retained after QC (dynamic 12% bad-channel threshold + spherical spline interpolation)

Methods

• Preprocessing: 1.0Hz highpass + 0.5-40Hz bandpass + pyprep PREP bad-channel detection
• Features: P300 mean amplitude (250-500ms window, 128 channels)
• Classifier: PCA(20) + LogisticRegression(C=1.0, balanced)
• Validation: Leave-One-Subject-Out CV + 1000-iteration permutation test

Project Structure

scripts/modma/
  run_modma_erp.py              # ERP pipeline (P300/N200, classification, ablation)
  modma_mdd_real_experiment.py  # Resting-state pipeline (QC, band power)
  run_modma_real.py             # Resting-state classification entry
  run_modma_reproduce.py        # Reproduction script
  modma_mdd_classification.py   # Early classification experiments
tests/                          # Unit tests
docs/PRE_REGISTRATION.md        # Cross-dataset replication pre-registration

Negative Results

Attempt	Result	Conclusion
P300+N200 enriched features (256-dim)	BA 0.670→0.554	Curse of dimensionality
Three-condition fusion (384-dim)	BA 0.670→0.521	Fusion introduces noise
Resting-state band power (640-dim)	BA=0.391	Below chance level
TDBRAIN cross-dataset replication	BA=0.500	Resting-state features don't generalize

Limitations

• N=52 limits generalizability
• No same-paradigm external validation dataset
• BA=0.670 is moderate; limited clinical applicability

Usage

pip install mne scikit-learn pyprep numpy scipy joblib pandas matplotlib
python scripts/modma/run_modma_erp.py

Download data from MODMA into the data/ directory first.

License

MIT