Human Stress Detection With Wearable Sensors Using Convolutional Neural Networks

In this article, we propose and evaluate a deep learning architecture based on convolutional neural networks for human stress detection using wearable sensors. This architecture has a first part that includes three convolutional layers for learning features from several biosignals. The second part is composed of three fully connected layers for stress detection. Additionally, we analyze several biosignal processing techniques to be applied before defining the inputs to the deep learning architecture: Fourier transform, cube root and constant Q transform. This analysis was performed on a public dataset, wearable stress and affect detection dataset (recorded by Robert Bosch GmbH Corporate Research, Germany), using a leave-one-subject-out cross-validation. We evaluated three different classification tasks including different emotional states. The accuracy increased from 93.1% to 96.6% when classifying stress vs. nonstress states, from 80.3% to 85.1% for differencing baseline versus stress versus amusement, and from 77.1% to 82.1% when discriminating five classes: baseline, stress, amusement, meditation, and recovery.