Stress is natural, especially during this unprecedented COVID-19 crisis. However, it can lead to serious health problems if experienced persistently over a long time. Studies conducted in the past have shown the significance of Electrodermal Activity (EDA) and Heart Rate Variability (HRV) in stress prediction. However, most of these previous works do not include personalization in real-time stress prediction. Therefore, we propose using incremental learning for continuous learning and personalization of the machine learning model that predicts stress arousal using HRV indices and EDA measurements collected constantly via a wearable device. Our system also provides personalized context-based recommendations for a stress-relieving activity when the user is found stressed. The recommended activity can be tried in our mobile application. Studies conducted in the past that aimed at providing such recommendations involved more user involvement, like filling questionnaires to get an idea of activity preferences at the beginning. We propose using Thompson Sampling instead of relying on user-filled questionnaires to determine the activity preferences in the cold start phase for the personalized recommendations. Once we get a fair idea of the effectiveness of activities for the given user, we switch to a clustering-based algorithm for providing recommendations. Thus, we present a digital solution that involves both stress prediction and mitigation with the help of a mobile application. We got an accuracy of 92% while pre-training the stress prediction model using the WESAD dataset, which further improved upon personalizing the model using incremental learning. The recommendations and activities supported by the application were found effective 71.5% of the times on average by our users.

Team Members:

Akshyta Katyal (IIIT Delhi), Anushika Gupta (IIIT Delhi)