PharmaHacks 2026 — EEG-based AD vs CN classification using a 3-model ensemble.
| Model | Features | LOOCV | Description |
|---|---|---|---|
| XGBoost Lasmar | 418 (Welch PSD @ 128Hz) | ~81% | Band power, ratios, Hjorth, spectral entropy |
| DICE-net | RBP + PLV tensors | TBD | Dual-branch CNN + Transformer, multi-seed |
| XGBoost v3 | 1330 (+ entropy + coherence) | TBD | Nonlinear features + Optuna tuning |
pip install -r requirements.txt
python run_all.pyThis runs all 3 models (LOOCV + final training + test predictions) and combines them via weighted ensemble. Final output: predictions.csv.
# XGBoost Lasmar
python models/xgboost_lasmar/train.py
# DICE-net (requires precomputed features)
python models/dicenet/precompute_features.py
python models/dicenet/predict.py
# XGBoost v3
python models/xgboost_v3/train.py
# Ensemble (after all models have run)
python ensemble.pytraining/AD/— 25 AD subjects (.npy, 19 channels)training/CN/— 28 CN subjects (13 labeled)training/test/— 14 unlabeled test subjectstraining/train_label_mapping.csv— subject labels