Mental Attention States Classification Using EEG Data

This project implements a mental state classification system using EEG data to detect three distinct cognitive states: focused, unfocused, and drowsy. The system utilizes both time-domain and frequency-domain approaches for feature extraction and classification.

Project Overview

The project implements multiple approaches to classify mental states:

• Time-domain analysis using 1D CNN-LSTM hybrid model
• Frequency-domain analysis using traditional ML approaches (Random Forest and SVM)
• Comprehensive signal processing pipeline including ICA for artifact removal

Dataset Description

The data was collected from 5 participants over 7 experimental sessions, with each session lasting 45-55 minutes. Participants operated a simulated train under controlled conditions:

• First 10 minutes: Focused state - actively controlling the simulation
• Next 10 minutes: Unfocused state - stopped monitoring but remained awake
• Final 10 minutes: Drowsy state - allowed to relax and doze

Data Collection Details

• 14 EEG channels: AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8, AF4
• Sampling rate: 128 Hz
• Channel locations cover frontal, temporal, parietal, and occipital regions
• Includes gyroscope data (GYROX, GYROY) for motion artifact detection

Methodology

Signal Processing Pipeline

1. Preprocessing

   • Bandpass filtering (0.3-30 Hz) to remove DC offset and high-frequency noise
   • ICA (Independent Component Analysis) for artifact removal:
     • Identifies and removes eye blinks, muscle artifacts, and cardiac signals
     • Uses FastICA algorithm with 14 components
     • Manual component selection based on topographic maps and time series
   • Motion artifact detection using gyroscope data

2. Feature Engineering

   a. Time Domain Approach:

   • Raw signal windows (30 seconds with 25% overlap)
   • Minimal preprocessing to preserve temporal patterns
   • Direct input to CNN-LSTM model

   b. Frequency Domain Approach:

   • Power spectral density estimation using Welch's method
   • Feature extraction from standard frequency bands:
     • Delta (0.5-4 Hz): Deep sleep indicators
     • Theta (4-8 Hz): Drowsiness and meditation
     • Alpha (8-13 Hz): Relaxed wakefulness
     • Beta (13-30 Hz): Active thinking and focus
   • Statistical features:
     • Mean power in each band
     • Peak frequency
     • Spectral entropy
     • Band power ratios

Classification Models

1. CNN-LSTM Hybrid (Time Domain):

   • 3 Conv1D layers with max pooling
   • 2 LSTM layers (128, 64 units)
   • Dropout layers for regularization
   • Softmax output for 3-class classification

2. Random Forest (Frequency Domain):

   • 1000 trees
   • Maximum depth: 10
   • Feature importance-based selection
   • Class weight balancing

3. SVM (Frequency Domain):

   • RBF kernel
   • Grid search for hyperparameter optimization
   • Feature selection using Random Forest importance scores

Results

Performance metrics across different models:

Model	Accuracy	Precision	Recall	F1-Score
Random Forest	0.81	0.81	0.81	0.81
CNN-LSTM	0.74	0.73	0.74	0.73
SVM (features selected)	0.73	0.74	0.73	0.73

Model Characteristics

• CNN-LSTM: Best at detecting state transitions and temporal patterns
• Random Forest: Excellent performance with frequency domain features
• SVM: Strong baseline performance, improved with feature selection

Key Findings

1. Frequency Domain Analysis:

   • Most discriminative features found in alpha and beta bands
   • Spectral power ratios highly effective for state discrimination
   • ICA crucial for improving signal quality

2. Time Domain Analysis:

   • CNN-LSTM effectively learns temporal dependencies
   • Direct signal processing reduces information loss
   • More robust to individual variations

3. Classification Performance:

   • Highest accuracy in distinguishing focused vs. drowsy states
   • Most challenging: differentiating unfocused from other states
   • Random Forest shows best overall performance

Future Work

1. Real-time Implementation:

   • Optimize processing pipeline for online analysis
   • Develop streaming data handling
   • Reduce classification latency

2. Model Improvements:

   • Investigate deep learning architectures for frequency domain
   • Develop hybrid feature extraction approaches
   • Implement attention mechanisms for temporal modeling

3. Applications:

   • Integration with attention monitoring systems
   • Driver drowsiness detection
   • Workplace safety monitoring

Contributors

• Vo Minh Thinh
• Nguyen Truong Thinh
• Tran Binh Phuong
• Nguyen Hong Son