https://github.com/niranjanrao07/adhd-ml-project

This project used machine learning to classify ADHD based on EEG data. We preprocessed the EEG signals, extracted various features, and used LDA for dimensionality reduction. A voting ensemble of classifiers achieved 72% accuracy in distinguishing between ADHD and control groups.
https://github.com/niranjanrao07/adhd-ml-project

adhd ensemble feature-engineering machine-learning preprocessing

Last synced: about 1 year ago
JSON representation

This project used machine learning to classify ADHD based on EEG data. We preprocessed the EEG signals, extracted various features, and used LDA for dimensionality reduction. A voting ensemble of classifiers achieved 72% accuracy in distinguishing between ADHD and control groups.

• Host: GitHub
• URL: https://github.com/niranjanrao07/adhd-ml-project
• Owner: NiranjanRao07
• Created: 2025-04-28T04:07:37.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2025-04-30T02:13:21.000Z (about 1 year ago)
• Last Synced: 2025-04-30T03:23:54.411Z (about 1 year ago)
• Topics: adhd, ensemble, feature-engineering, machine-learning, preprocessing
• Language: Jupyter Notebook
• Homepage:
• Size: 1.35 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# ADHD Detection from EEG Data

This repository contains all code, data, and artifacts for our project on automated ADHD detection using 19-channel EEG recordings and classical machine learning techniques.

## 🔍 Overview

1. **Data Exploration**

   - MATLAB `.mat` EEG files (61 ADHD, 60 control), 19 channels, 128 Hz sampling

   - Scripts compute channel-wise stats: mean, variance, skewness, kurtosis, ptp

2. **Preprocessing**

   - Butterworth band-pass filter (0.5–45 Hz) removes drift & noise

   - Z-score normalization per channel (mean = 0, std = 1)

   - Amplitude-based artifact removal (±100 µV threshold)

3. **Feature Extraction**

   - **Time-domain:** mean, std, skewness, kurtosis, RMS, zero crossings, peak-to-peak

   - **Frequency-domain:** Welch PSD, delta/theta/alpha/beta/gamma band powers, spectral entropy, SEF, PSD slope

   - **Non-linear:** approximate entropy, Higuchi fractal dimension, Hjorth mobility/complexity, Hurst exponent

4. **Dimensionality Reduction**

   - PCA (95 % variance) tested → suboptimal

   - **Final:** supervised LDA applied only on training split → single discriminant axis

5. **Modeling & Evaluation**

   - Classifiers on LDA output: SVM, Decision Tree, Random Forest, KNN, Logistic Regression

   - Hyperparameter tuning: grid search + 5-fold CV

   - Held-out 20 % test split for final metrics

   - Voting ensemble of all five models

6. **Results**

   - **Ensemble Test Performance:**

     - Accuracy: 72.0 %

     - Precision: 68.8 %

     - Recall: 84.6 %

     - F1-Score: 75.9 %

     - ROC AUC: 78.8 %

Explore raw data: run `analyze_files.py`

Preprocess & extract features: open and execute `data_preprocess.ipynb` and `feature_extraction.ipynb`

Train & evaluate models: implement LDA on train split, train classifiers, build ensemble

## 🤝 Collaboration & Version Control

Progress and deliverables were tracked through regular team syncs and shared document updates.