Stress Classification using single-channel ECG

VR Stress Assesment from proprietary dataset

Main Approaches

Note: Using Leave-One-Subject-Out (LOSO) evaluation paradigm for all

Models

1. InceptionTime architecture (source repo)

   • Regression (224x1 time-series input to 224x1 labels)
   • Original: Multiclass classification (224x1 input to 3-class output)

2. SE-Net variants

   • SE-ResNet14
   • ISE-ResNet14
   • ISE-Net matching Multi-ECG Classification paper

Data Processing

• Time-series window: 1 second, 5 second, or 224-samples (1 R-R cycle)
• Post processing: Gramien Angular Fields (GAF), Markov Transition Fields (MTF), Recurrence Plots (RP)

Files

Jupyter notebooks

• data_exploration.ipynb - data exploration and statistics
• inceptiontime_ecg.ipynb - LOSO experiments with InceptionTime
• se-net_3class.ipynb - LOSO experiments with SENets

Other

• InceptionTime.py - InceptionTime model

Visualization

• results_visualization.ipynb - Visualizing training and validation accuracy and loss across different LOSO users with different preprocessing techniques. Can be adapted for general training visualization too! Sample plot seen below:

References

• InceptionTime - https://github.com/hfawaz/InceptionTime