lebiraja/eeg-depression-v4

EEG-Based Depression (MDD) Detection — V4 Ensemble

Leave-One-Subject-Out (LOSO) cross-validated EEG classifier for Major Depressive Disorder, achieving 96.88 % subject-level accuracy and 99.80 % AUC-ROC on 64 subjects from the public figshare EEG dataset.

---

Model Architecture

This is a 3-model heterogeneous ensemble:

Component	Details
1D-CNN (weight=0.60)	Multi-scale convolution stem → 4 × ResBlock1D + SEBlock1D → GlobalAvgPool → MLP. Input: raw EEG (19 ch × 1000 samples).
XGBoost (weight=0.25)	800 estimators, max_depth=6, trained on 1047-dim handcrafted features.
SVM (weight=0.15)	RBF kernel (C=10, γ=scale), trained on the same 1047-dim features.

Epoch-level probabilities from each model are aggregated as a weighted sum, then subject-level probability is the trimmed mean (5 % trim) of epoch probabilities.

Handcrafted Feature Set (1047 dimensions)

Group	Dims	Description
Spectral power	95	Band power per channel (δ/θ/α/β/γ)
Temporal stats	171	Mean, variance, skewness, kurtosis, Hjorth params, zero-crossing rate
Wavelet / WPD	304	Wavelet Packet Decomposition energy + entropy
Connectivity	342	Phase Locking Value + coherence between all channel pairs
Asymmetry	40	Frontal/temporal α asymmetry (8 electrode pairs)
Band ratios	95	Cross-band power ratios per channel

---

Performance (LOSO Cross-Validation, 64 subjects)

Metric	Threshold=0.50	Threshold=0.575
Subject Accuracy	90.62 % (58/64)	96.88 % (62/64)
AUC-ROC	99.80 %	99.80 %
Sensitivity (MDD recall)	100 %	100 %
Specificity (Healthy recall)	80.0 %	93.3 %
F1 Score	93.15 %	97.14 %

Post-hoc threshold optimisation on LOSO predictions (threshold=0.575) yields 96.88 % accuracy while maintaining 100 % sensitivity — no MDD subject is ever missed.

---

Dataset

• Source: figshare EEG dataset
• Subjects: 64 (34 MDD, 30 Healthy Controls)
• EDF files: 181
• Channels: 19 (standard 10-20 system)
• Signal: Bandpass 1–45 Hz, notch 50/60 Hz, resampled to 250 Hz
• Epochs: 4 s, 50 % overlap → ~570 epochs/subject on average
• Total epochs: 36,247

---

Repository Files

models/
  final/
    cnn_final.pt                  # CNN state dict (trained on all 64 subjects)
    xgboost_final.json            # XGBoost model (all 64 subjects)
    svm_and_scaler_final.pkl      # SVM + StandardScaler (all 64 subjects)
  fold_NN_SID/                    # One directory per LOSO fold (64 total)
    cnn_weights.pt
    xgboost.json
    svm_and_scaler.pkl
checkpoints/
  fold_NN_SID.pkl                 # Per-fold LOSO results (probs, labels, metrics)
config/
  model_config.yaml
  data_config.yaml
  training_config.yaml
results_best.json                 # Full LOSO metrics from training
test_loso_results.json            # Metrics from test_loso.py

---

Usage

Load CNN

import torch
from eeg_depression_detection.models.full_model import EEG1DCNN  # adjust import path

CNN_CONFIG = dict(
    n_channels=19,
    n_samples=1000,
    n_classes=2,
)
model = EEG1DCNN(**CNN_CONFIG)
state = torch.load("models/final/cnn_final.pt", map_location="cpu")
model.load_state_dict(state)
model.eval()

Load XGBoost

import xgboost as xgb
xgb_model = xgb.XGBClassifier()
xgb_model.load_model("models/final/xgboost_final.json")

Load SVM + Scaler

import pickle
with open("models/final/svm_and_scaler_final.pkl", "rb") as f:
    bundle = pickle.load(f)
svm_model = bundle["svm"]
scaler    = bundle["scaler"]

---

Citation

If you use this model or dataset pipeline in your research, please cite:

@misc{eeg_depression_v4,
  author       = {Pranav},
  title        = {EEG Depression Detection V4 Ensemble},
  year         = {2026},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/lebiraja/eeg-depression-v4}},
}

---

Licence

MIT — see LICENSE

Generated 2026-03-11