Predicting Stress in Older Adults with RNN and LSTM from Time Series Sensor Data and Cortisol

We explore the problem of stress prediction with a combination of behavioral time series sensor data, including Electrodermal Activity (EDA), Blood Volume Pulse (BVP), and Skin Temperature (ST) employing Recurrent Neural Networks (RNNs). The dataset is labeled with cortisol biomarker and TSST protocol for binary stress levels. Leveraging the capabilities of RNNs, specifically Long Short-Term Memory (LSTM) net-works, this research examines the relationship between various behavioral patterns captured by sensors and cortisol levels. The experiment includes data from a targeted age group of 40 older adults. The RNN models are trained to forecast stress levels up to 4.5 minutes ahead, enabling proactive intervention strategies. We report an accuracy of 94% and 96% for RNN and LSTM respectively. Our early investigation showcases the potential of integrating multiple data modalities for accurate stress prediction, underscoring the importance of time series machine learning techniques in stress monitoring applications.