🧠 EEG Burnout Classification

🔍 Project Overview

This project aims to develop a machine learning pipeline to classify burnout levels—Low, Medium, or High—based on raw EEG signal data. The workflow involves preprocessing EEG recordings, segmenting them into meaningful chunks, extracting statistical features, and training classifiers to accurately predict burnout levels.

---

📁 Dataset Description

• Source: Raw EEG data (~3GB) collected from 30 patients.
• Conditions: Eyes Open & Eyes Closed.
• Format:
  • Each patient has multiple .csv files (up to 3 per condition).
  • Each file contains ~3 minutes of EEG data (~50,000 rows).
• Labels: Each EEG file is labeled using Maslach Burnout Inventory (MBI) scores.

---

🧹 Data Preprocessing

✅ Cleaning

• Removed null and duplicate values.
• Verified timestamp and signal consistency.

✅ Labelling

• Each EEG file is associated with a burnout level from the subject’s MBI score.

✅ Segmentation

• Method: Sliding window.
• Window Size: 10 seconds (to avoid data loss from sample-count-based segmentation).
• Outcome: ~18 segments per 3-minute file.

---

🧠 Feature Extraction

From each 10-second EEG segment, features are extracted from four primary channels:

• RAW_AF7, RAW_AF8, RAW_TP9, RAW_TP10

Extracted Features:

• Statistical: Mean, Std Deviation, Skewness, Kurtosis
• Signal-based: Zero Crossing Rate, Line Length
• Hjorth Parameters: Activity, Mobility, Complexity
• Demographic: Encoded Gender

💡 A total of 36 features are generated per segment.

---

🤖 Model Training & Evaluation

1️⃣ Logistic Regression (L2 Regularized)

• Accuracy: 70%
• Cross-Validation (5-fold): 69.43%
• Precision (macro): 93%
• F1-score (macro): 92%
• Recall (macro): 92%

⚠️ Good metrics but poor generalization—overfit on small data (~1GB), degraded on full dataset (~3GB).

---

2️⃣ Random Forest Classifier (Final Model)

• Accuracy: 92.4%
• Cross-Validation (5-fold): 92.26%
• Precision (macro): 93%
• F1-score (macro): 92%
• Recall (macro): 92%

✅ Superior generalization, robust to non-linear patterns, performs consistently across datasets.

---

🧩 Challenges & Solutions

Challenge	Solution
Data loss due to fixed-size segments	Switched to time-based (10s) sliding window segmentation
Poor generalization in simple models	Adopted Random Forest for handling complex EEG variations
Large data volume (3GB)	Applied memory-efficient processing and optimized loops
Label inconsistencies	Ensured each file is labeled using subject's unique MBI-derived burnout

---

🚀 Deployment

The final model is integrated into a Streamlit-based GUI, allowing:

• Upload of raw EEG .csv files.
• Automated processing & feature extraction.
• Burnout level prediction.
• Personalized advisory message based on predicted state.

🎯 Designed for non-technical users like clinicians or researchers for easy mental health assessment using EEG data.

---

✅ Conclusion

This project demonstrates how feature-based machine learning can effectively classify burnout levels using EEG signals. With systematic preprocessing, robust feature extraction, and model evaluation, this approach shows promise for real-world mental health diagnostics.

---

🛠️ Technologies Used

• Python
• Pandas, NumPy
• Scikit-learn
• Streamlit
• SciPy
• EEG signal processing techniques

---

📌 Future Work

• ✅ GUI Deployment (Completed)
• ⏳ Model Fine-tuning with additional datasets
• ⏳ Integration with wearable EEG devices

---

📬 Contact

For questions or collaborations, feel free to reach out