Skip to content

xyjk0511/tdbrain-eeg-classification

BranchesTags

Repository files navigation

TDBRAIN EEG Classification / TDBRAIN 脑电分类项目

Python Dataset AUC Subjects

At a glance / 项目速览

Item Summary
Task MDD vs ADHD EEG classification / 抑郁症与 ADHD 脑电分类
Dataset TDBRAIN
Best result AUC 0.798
Validation 5-fold StratifiedGroupKFold + permutation test
Project value EEG ML, feature engineering, strict leakage control

TDBRAIN summary card

Visual snapshot / 可视化结果

ROC Permutation Test
Ensemble ROC Permutation null distribution

Resting-state EEG classification pipeline for distinguishing psychiatric conditions on the TDBRAIN dataset, with subject-level validation, feature engineering, and statistically validated performance.

一个基于 TDBRAIN 数据集的静息态脑电分类项目，重点在于受试者级验证、特征工程，以及经过统计检验的模型表现。

Overview / 项目概述

Goal / 目标

  • Classify Major Depressive Disorder (MDD) versus ADHD from resting-state EEG.
  • 使用静息态 EEG 区分 抑郁症（MDD） 与 注意缺陷多动障碍（ADHD）。

This repository represents a practical EEG ML workflow rather than a toy example.

这个仓库体现的是一个真实 EEG 机器学习工作流，而不是简单的示例代码。

Results / 核心结果

Model AUC Accuracy Sensitivity Specificity p-value
SVM 0.770 0.720 0.797 0.596 < 0.001
Random Forest 0.796 0.765 0.862 0.606 < 0.001
XGBoost 0.796 0.753 0.823 0.638 < 0.001
Ensemble (soft-vote) 0.798 — — — —

Dataset summary / 数据概况

  • 493 subjects (305 MDD, 188 ADHD)
  • dual-condition EEG: eyes-open + eyes-closed
  • strict 5-fold StratifiedGroupKFold subject isolation
  • 1000-permutation test, all main models with p < 0.001

Feature Design / 特征设计

The pipeline uses 992 dimensions of EEG-derived features, combining both eyes-open and eyes-closed conditions.

该项目使用 992 维 EEG 特征，融合了睁眼与闭眼两种静息态条件。

Feature families / 特征类别

  • absolute band power
  • relative band power
  • TBR + FAA
  • Hjorth parameters
  • broadband spectral entropy
  • per-band spectral entropy

Why this matters / 为什么这很重要

Instead of using a fully black-box model, the project emphasizes interpretable signal features with links to EEG literature.

它没有直接依赖黑箱模型，而是强调与 EEG 文献相呼应、可解释的特征设计。

Pipeline / 技术流程

Raw EEG (.bdf)
  → rereference + filtering
  → epoching + artifact rejection
  → feature extraction
  → EO + EC concatenation
  → train/validation split with subject isolation
  → model selection and evaluation
  → permutation test and interpretation

Key modeling choices / 关键建模设计

  • group-aware cross-validation to prevent subject leakage
  • oversampling and feature selection inside the training pipeline
  • threshold optimization via Youden index
  • statistical significance testing via permutation

Repository Structure / 仓库结构

config.py                  # dataset paths, EEG channels, constants
preprocessor.py            # filtering, rereferencing, epoching, artifact rejection
feature_extractor.py       # PSD, Hjorth, entropy features
classifier.py              # model training, validation, ensemble, permutation test
data_loader.py             # raw EEG loading
connectivity_extractor.py  # related connectivity features / extensions
main.py                    # end-to-end pipeline entry
results.json               # result summary
shap_summary.json          # SHAP-based interpretation summary
tests/                     # tests

Why this project matters / 为什么这个项目重要

This is one of the clearest examples of how I approach biomedical ML problems:

  • domain-aware feature engineering
  • strict leakage control
  • statistically responsible evaluation
  • balancing interpretability and performance

这个项目能很好体现我做生物医学 ML 的方式：

  • 领域驱动的特征工程
  • 严格避免泄漏
  • 更负责任的统计验证
  • 在可解释性与性能之间找平衡

Reproducibility / 复现说明

Requirements

  • Python 3.11+
  • access to the TDBRAIN dataset

Install

pip install mne numpy scipy scikit-learn xgboost imbalanced-learn shap antropy mne-connectivity

Configure dataset path

DATASET_ROOT = Path("/path/to/tdbrain_dataset")

Run

python main.py

Outputs:

  • results.json
  • shap_summary.json

Notes / 说明

This repository is intended as a research-oriented project repository.

本仓库更适合作为研究型项目仓库来理解。

A future cleanup would likely include:

  • cleaner experiment configuration
  • clearer data preparation docs
  • better visualization of results and feature importance

后续如果继续整理，我会优先补充：

  • 更清晰的实验配置入口
  • 更明确的数据准备说明
  • 更完整的结果图和解释性可视化

Overview / 项目概述

Goal / 目标

  • Classify Major Depressive Disorder (MDD) versus ADHD from resting-state EEG.
  • 使用静息态 EEG 区分 抑郁症（MDD） 与 注意缺陷多动障碍（ADHD）。

This repository represents a practical EEG ML workflow rather than a toy example.

这个仓库体现的是一个真实 EEG 机器学习工作流，而不是简单的示例代码。

Results / 核心结果

Model AUC Accuracy Sensitivity Specificity p-value
SVM 0.770 0.720 0.797 0.596 < 0.001
Random Forest 0.796 0.765 0.862 0.606 < 0.001
XGBoost 0.796 0.753 0.823 0.638 < 0.001
Ensemble (soft-vote) 0.798 — — — —

Dataset summary / 数据概况

  • 493 subjects (305 MDD, 188 ADHD)
  • dual-condition EEG: eyes-open + eyes-closed
  • strict 5-fold StratifiedGroupKFold subject isolation
  • 1000-permutation test, all main models with p < 0.001

Feature Design / 特征设计

The pipeline uses 992 dimensions of EEG-derived features, combining both eyes-open and eyes-closed conditions.

该项目使用 992 维 EEG 特征，融合了睁眼与闭眼两种静息态条件。

Feature families / 特征类别

  • absolute band power
  • relative band power
  • TBR + FAA
  • Hjorth parameters
  • broadband spectral entropy
  • per-band spectral entropy

Why this matters / 为什么这很重要

Instead of using a fully black-box model, the project emphasizes interpretable signal features with links to EEG literature.

它没有直接依赖黑箱模型，而是强调与 EEG 文献相呼应、可解释的特征设计。

Pipeline / 技术流程

Raw EEG (.bdf)
  → rereference + filtering
  → epoching + artifact rejection
  → feature extraction
  → EO + EC concatenation
  → train/validation split with subject isolation
  → model selection and evaluation
  → permutation test and interpretation

Key modeling choices / 关键建模设计

  • group-aware cross-validation to prevent subject leakage
  • oversampling and feature selection inside the training pipeline
  • threshold optimization via Youden index
  • statistical significance testing via permutation

Repository Structure / 仓库结构

config.py                  # dataset paths, EEG channels, constants
preprocessor.py            # filtering, rereferencing, epoching, artifact rejection
feature_extractor.py       # PSD, Hjorth, entropy features
classifier.py              # model training, validation, ensemble, permutation test
data_loader.py             # raw EEG loading
connectivity_extractor.py  # related connectivity features / extensions
main.py                    # end-to-end pipeline entry
results.json               # result summary
shap_summary.json          # SHAP-based interpretation summary
tests/                     # tests

Why this project matters / 为什么这个项目重要

This is one of the clearest examples of how I approach biomedical ML problems:

  • domain-aware feature engineering
  • strict leakage control
  • statistically responsible evaluation
  • balancing interpretability and performance

这个项目能很好体现我做生物医学 ML 的方式：

  • 领域驱动的特征工程
  • 严格避免泄漏
  • 更负责任的统计验证
  • 在可解释性与性能之间找平衡

Reproducibility / 复现说明

Requirements

  • Python 3.11+
  • access to the TDBRAIN dataset

Install

pip install mne numpy scipy scikit-learn xgboost imbalanced-learn shap antropy mne-connectivity

Configure dataset path

DATASET_ROOT = Path("/path/to/tdbrain_dataset")

Run

python main.py

Outputs:

  • results.json
  • shap_summary.json

Notes / 说明

This repository is intended as a research-oriented project repository.

本仓库更适合作为研究型项目仓库来理解。

A future cleanup would likely include:

  • cleaner experiment configuration
  • clearer data preparation docs
  • better visualization of results and feature importance

后续如果继续整理，我会优先补充：

  • 更清晰的实验配置入口
  • 更明确的数据准备说明
  • 更完整的结果图和解释性可视化

About

EEG classification pipeline on TDBRAIN for psychiatric disorder prediction with cross-validation, feature engineering, and AUC 0.798

Topics

machine-learning neuroscience eeg feature-engineering mental-health psychiatric-disorders

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages