EEG Biomarkers for PTSD Diagnosis using Explainable AI (XAI)

An interpretable machine learning framework for identifying neurobiological markers of PTSD using EEG.

This repository implements a clinically transparent and computationally efficient approach for PTSD diagnosis using EEG, Machine Learning, and Explainable AI. Visual + methodological details are based on the research poster included in the repo.

---

Overview

Post-Traumatic Stress Disorder (PTSD) is currently diagnosed primarily through clinical interviews, which can be subjective and vary across clinicians. EEG provides a non-invasive, scalable, and cost-effective way to capture neural dynamics, but raw EEG is high-dimensional and difficult to interpret.

This project tackles these challenges by:

• Reducing high-dimensional EEG data to the most relevant features using ElasticNet
• Training multiple ML classifiers on these features
• Applying Explainable AI (XAI) tools to highlight interpretable neural biomarkers
• Identifying EEG frequency-band and connectivity patterns linked to PTSD

---

Dataset

Property	Description
Total Subjects	104 (52 PTSD, 52 Healthy)
Data Type	qEEG Resting-State
Features	1140 PSD & Functional Connectivity scores across 19 channels and 6 frequency bands
Preprocessing	Missing-value imputation + categorical encoding

---

Visualization of EEG Montages

The following image illustrates the EEG topographic maps for two conditions: Healthy Control and Posttraumatic Stress Disorder. These maps provide a visual comparison of EEG activity across different frequency bands.

---

Methodology

1. Preprocessing

   • Remove irrelevant fields
   • Impute missing values
   • Encode categorical variables

2. Feature Selection

   • ElasticNet regression reduces ~97% of features
   • Retains only clinically meaningful and interpretable EEG features

   ElasticNet Feature Weights

   The ElasticNet regression model was used to identify the most relevant EEG features by balancing L1 and L2 regularization. This approach ensures sparsity while retaining correlated features.

3. Model Training

   • ML models tested: SVC, KNN, RandomForest, XGBoost, LGBM, AdaBoost, CatBoost
   • Hyperparameter tuning performed using Optuna

   Model Feature Importance

   Feature importance was derived from the best-performing models to understand the contribution of each feature to PTSD classification.

4. Explainable AI

   • SHAP → Global interpretability (biomarker patterns)
   • LIME → Individual patient-level interpretation
   • ELI5 → Feature ranking & verification

   SHAP Summary Plot

   SHAP (SHapley Additive exPlanations) was used to visualize the global impact of features on model predictions. This plot highlights the most influential features across the dataset.

   Permutation Importance

   Permutation importance was calculated to validate the robustness of feature rankings and their impact on model performance.

---

Best Model Performance

Model	AUC Score	Notes
Random Forest (Optuna-Tuned)	0.85	Most stable & interpretable
CatBoost	0.836	Competitive performance
XGBoost	0.821	Robust but slightly less interpretable
LGBM	0.810	Efficient but sensitive to hyperparameters

Permutation testing confirmed significance at p < 0.05.

---

Key Findings

• Beta Power in Right Primary Motor Cortex (C4 region) emerged as a strong marker.

• Alpha band functional connectivity between frontal and parietal regions showed meaningful group differences.

• IQ scores contributed significantly and interactively to PTSD classification.

• PTSD patients show:

  • Elevated Beta / High-Beta (hyperarousal signature)
  • Reduced Delta / Alpha (attenuated cognitive & emotional regulation)

---

Repository Structure

├── .git/                   # Git version control metadata
├── .gitignore              # Git ignore rules
├── .vscode/                # VS Code workspace settings
├── README.md               # Project documentation
├── catboost_info/          # CatBoost training logs
│   ├── catboost_training.json
│   ├── learn/
│   ├── learn_error.tsv
│   └── time_left.tsv
├── data/                   # EEG dataset (not included due to privacy)
│   └── EEG_data.csv
├── main.py                 # Entry point for the project
├── notebooks/              # Jupyter notebooks for analysis
│   ├── prelim.ipynb
│   ├── visuals.ipynb
│   └── xai.ipynb
├── pipeline/               # Core pipeline scripts
│   ├── config.py
│   ├── data_loader.py
│   ├── evaluation.py
│   ├── feature_selection.py
│   ├── model_registry.py
│   ├── train_and_evaluate.py
│   └── xai.py
├── requirements.txt        # Python dependencies
├── results/                # Model outputs, SHAP plots, etc.
│   ├── bayes_results.csv
│   ├── elasticnet_weights.png
│   ├── final_auc.txt
│   ├── final_summary.csv
│   ├── model_feature_importance.png
│   ├── optuna_results.csv
│   ├── permutation_importance.png
│   ├── selected_features.txt
│   ├── shap_summary_plot.png
│   └── summary.png
├── src/                    # Source code for experiments
│   ├── bayes_optimization.py
│   ├── config.py
│   ├── grid_search.py
│   ├── optuna_optimization.py
│   ├── preprocessing.py
│   ├── train_and_evaluate.py
│   └── xai_interpretation.py

---

Future Work

• Multi-disorder EEG classification (PTSD vs MDD vs GAD)
• Cross-session and cross-device generalization
• Clinical tool: subject-level risk dashboard using SHAP/LIME profiles
• Multimodal fusion with ECG, behavioral, or speech biomarkers

---

Contributors

Lead Researchers

• ParmarDarshan29
• twi-exe

Supervision

• Dr. Shankar Parmar