AI Projects

Dataset Overview

• Subjects: EEG signals were recorded from 72 subjects.

• Recording Duration: Each recording lasted 60 seconds, captured both before and during arithmetic task performance.

• Segmentation: The 60-second recordings were divided into 30 segments, each lasting 2 seconds.

EEG Data and Feature Extraction

• Channels: EEG data was collected from 21 channels.

• Features: A total of 10 features were extracted from each channel, resulting in 210 features per segment. The extracted features fall into three categories:

  • Bandpower Features (6 features):

    1. Delta Bandpower

    2. Theta Bandpower

    3. Alpha Bandpower

    4. Sigma Bandpower

    5. Beta Bandpower

    6. Gamma Bandpower

  • Fractal Dimension Features (2 features):

    1. Petrosian Fractal Dimension

    2. Higuchi Fractal Dimension

  • Entropy-Based Features (2 features):

    1. Sample Entropy

    2. Permutation Entropy

Classification Task

Seven machine learning models, including an artificial neural network (ANN), were trained and evaluated on the dataset. The dataset was split into 80% for training and 20% for testing. Below are the performance results of the individual models on the test data:

• Decision Tree: 86.74% accuracy

• Support Vector Classifier: 85.25% accuracy

• Logistic Regression: 83.76% accuracy

• k-Nearest Neighbors (kNN): 96.89% accuracy

• XGBoost: 98.38% accuracy

• Random Forest: 96.08% accuracy

• Artificial Neural Network (ANN): 99.05% accuracy

Ensemble Technique

To further enhance classification performance, an ensemble technique was implemented by combining the top three models: kNN, XGBoost, and ANN. This ensemble approach achieved remarkable results across all performance metrics:

• Precision: 99.46%

• Recall: 99.46%

• F1-Score: 99.46%

• Accuracy: 99.46%

Summary

The analysis demonstrates the exceptional capability of machine learning and ensemble techniques in classifying mental stress from EEG-signals while performing arithmetic task. The ANN model individually achieved the highest accuracy (99.05%), while the ensemble model outperformed all individual models achieving a accuracy of 99.46%, along with near-perfect performance across all metrics.