Deep Learning Based Stress Assessment Using PPG Signals from WESAD Dataset

Srejita Chakraborty, Rashmi Kumari, Pabitra Das, Surita Sarkar, Souris Sahu, Saurabh Pal, Amit Acharyya

Abstract: Owing to the increased prevalence of stress in modern society along with its adverse effects on human wellbeing, early intervention becomes crucial for ensuring a healthier life. Stress can be the cause of various medical abnormalities, like, heart disorders, raised blood pressure, diabetes, and so on. Considering the critical impact of stress, numerous research studies are being conducted to accurately predict stress levels using physiological signals such as Photoplethysmogram (PPG), Electrocardiogram (ECG), Electrodermal Activity (EDA), Electromyography (EMG) and Respiration (Resp), which, in turn, may lead to an increased requirement for sensors. In this regard, PPG based stress analysis has gained popularity among researchers due to its easy, cost-effective and non-invasive acquisition protocol along with its wide usage in wrist-worn wearable devices like smartwatches. This work presents a novel stress detection technique using raw PPG signal from publicly available WESAD (Wearable Stress and Affect Detection) dataset. Short 5sec segments of pre-processed PPG signals were fed to a CNN-LSTM model for detection of stressful events. The model achieved an accuracy of 96.90%, precision of 95.96%, recall of 97.65% and F1 score of 96.7% on the test data showcasing an improvement over other reported works by $\approx 5 \%$ in accuracy.

External IDs:dblp:conf/newcas/ChakrabortyKDSSPA25