ADHD EEG Classification

Classifying ADHD vs. healthy children from resting-state EEG using handcrafted spectral and statistical features with classical machine learning.

A subject-level analysis on the IEEE 19-channel ADHD EEG dataset (Nasrabadi et al., 2020), implemented with scipy, scikit-learn, and pandas.

→ View the interactive research showcase

Background

Attention Deficit Hyperactivity Disorder (ADHD) is one of the most common neurodevelopmental disorders in children. EEG-based methods are an active area of research as a non-invasive, objective complement to behavioral diagnostic criteria.

Two findings from the literature motivate this project:

1. The theta/beta ratio (TBR) has historically been proposed as an ADHD biomarker; elevated theta and reduced beta activity over frontal-central regions.
2. Multi-feature ML approaches consistently outperform any single biomarker, suggesting ADHD-related EEG signatures are distributed across bands and channels.

This project builds a clean, end-to-end pipeline to test these ideas on a public benchmark dataset.

---

Dataset

IEEE EEG Data for ADHD / Control Children (Nasrabadi et al., 2020) - IEEE DataPort | Kaggle mirror

Property	Value
Subjects	121 (61 ADHD, 60 Control)
Age range	7–12 years
Channels	19 (10–20 standard)
Sampling rate	128 Hz
Task	Visual attention (cartoon character counting)
Mean recording duration	154s ADHD / 125s Control

ADHD subjects were diagnosed by an experienced psychiatrist according to DSM-IV criteria. Controls had no history of psychiatric disorders, epilepsy, or high-risk behaviors.

---

Methodology

1. Exploratory Data Analysis (01_eda.ipynb)

• Class balance and recording duration analysis
• Raw EEG signal inspection; ADHD signals appear noisier/more irregular than Control
• Power Spectral Density (PSD) comparison using Welch's method
• Group-level band power comparison across all 121 subjects

2. Feature Extraction (02_feature_extraction.ipynb)

171 features extracted per subject:

Feature group	Count	Description
Band power	76	Delta (1–4 Hz), Theta (4–8), Alpha (8–12), Beta (12–30) × 19 channels
Theta/Beta ratio	19	One TBR per channel
Statistical	76	Mean, Variance, Skewness, Kurtosis × 19 channels

Band power computed using Welch's method (nperseg=256), which is more stable than raw FFT on noisy, variable-length EEG.

3. Classification (03_classification.ipynb)

• Models: Random Forest (200 trees, min_samples_leaf=2), SVM with RBF kernel (C=10, gamma='scale')
• Preprocessing: StandardScaler inside an sklearn Pipeline (critical for SVM)
• Evaluation: Stratified 5-fold cross-validation
• Out-of-fold predictions used for confusion matrices and ROC curves to avoid data leakage

---

Results

Cross-validated performance (5-fold)

Model	Accuracy	F1 Score	ROC-AUC
Random Forest	0.686 ± 0.042	0.700 ± 0.053	0.754 ± 0.048
SVM (RBF)	0.654 ± 0.062	0.651 ± 0.039	0.683 ± 0.085

Random Forest outperforms SVM across all three metrics. The AUC of 0.75 indicates moderate but meaningful separability which is well above chance (0.5), but lower than the ~85–95% accuracy figures sometimes reported in the literature. Those higher numbers typically rely on segmenting recordings into thousands of short windows (which inflates sample count and risks within-subject leakage); the evaluation here is subject-level, which is the honest and clinically meaningful benchmark.

Most important features (Random Forest)

The top 20 features were dominated by:

• Signal variance over parietal/temporal regions - Var_P7, Var_P3, Var_T7, Var_Pz, Var_F7, Var_O2, Var_Fz - the strongest single signal in the data
• Theta/Beta Ratio at the parietal midline (TBR_Pz) - ranked #3, supporting the clinical biomarker hypothesis at a posterior site rather than the more commonly studied frontal-central one
• Alpha activity over central midline (Alpha_Cz) - ranked #4
• Beta band activity across central and frontal regions (Beta_Cz, Beta_F4, Beta_C3, Beta_Fp2)

The takeaway: ADHD-related EEG differences in this dataset are distributed across signal-variability and spectral features, not localized to one band or biomarker. Models built around a single feature (such as raw theta/beta ratio) under-perform multi-feature models, which is consistent with the literature.

Theta/Beta ratio: the conventional biomarker

The mean TBR (averaged across all channels) was actually higher in the Control group (2.99) than in ADHD (2.54) in this dataset. This is opposite to the classical "elevated TBR in ADHD" finding, but it is consistent with several recent studies questioning TBR as a standalone diagnostic. TBR still contributed predictive value when used per-channel (especially at Pz), but it is not a clean separator on its own. This motivated the multi-feature approach used here.

---

Project Structure

adhd-eeg-classification/
├── data/
│   └── adhdata.csv                  # Download from Kaggle (not committed)
├── notebooks/
│   ├── 01_eda.ipynb                 # Exploratory data analysis
│   ├── 02_feature_extraction.ipynb  # Feature engineering (produces features.csv)
│   ├── 03_classification.ipynb      # Model training and evaluation
│   └── 04_xai.ipynb                 # SHAP explainability and site data export
├── website/                         # React research showcase site
├── requirements.txt
└── README.md

---

How to Reproduce

1. Clone the repository

   git clone https://github.com/evasafi/adhd-eeg-classification.git
   cd adhd-eeg-classification

2. Install dependencies

   pip install -r requirements.txt

3. Download the dataset from Kaggle and place adhdata.csv in data/

4. Update the file path at the top of each notebook to point to your local data/adhdata.csv

5. Run the notebooks in order:

   • 01_eda.ipynb — exploratory analysis (~3 minutes)
   • 02_feature_extraction.ipynb — produces features.csv (~3 minutes)
   • 03_classification.ipynb — model training and evaluation (~1 minute)

6. Run 04_xai.ipynb — computes SHAP values, generates topographic maps, exports site_data.json for the website

---

Limitations

• Sample size. N=121 subjects is modest for deep learning; results may not generalize.
• No artifact removal. Eye blinks, muscle artifacts, and electrode drift were not removed. A production pipeline would use ICA or autoreject.
• Single dataset. Generalization to other recording setups, age ranges, or clinical populations is untested.
• Subject-level features only. Time-windowed or connectivity-based features could potentially improve performance.
• Not for clinical use. This is a research/educational project, not a diagnostic tool.

---

References

• Nasrabadi, A. M. et al. (2020). EEG data for ADHD / Control children. IEEE DataPort.
• Snyder, S. M. et al. (2015). Integration of an EEG biomarker with a clinician's ADHD evaluation. Brain and Behavior.
• Multiple papers on this dataset are tracked at PapersWithCode.

---

License

MIT