Transfer Learning for Subject-Independent Stress Detection using Physiological Signals

The main objective of this work is to investigate the role of transfer learning for enhancing the performance of subject-independent stress detection models. Using Time Series Feature Extraction on basis of Scalable Hypothesis tests (tsfresh), features are extracted from raw Electrocardiogram (ECG) and Electrodermal Activity (EDA) data and Convolutional Neural Network (CNN) models are built for stress detection. We used benchmark datasets-ASCERTAIN and CLAS to conduct our experiments. The performance of CNN models (subject-independent models) are compared with and without transfer learning. Results show a 1-4 percentage (absolute) increase in transfer learning, so these results serves as the basis for future studies in this direction.