Multilevel mental stress detection using ultra-short pulse rate variability series

Abstract

Prolonged exposure to mental stress reduces human work efficiency in daily life and may increase the risk of diabetes and cardiovascular diseases. However, identification of the true degree of stress in its initial stage can reduce the risk of life threatening diseases. In this paper, we proposed a multilevel stress detection system using ultra-short term recordings of a low cost wearable sensor. We designed an experimental paradigm based on Mental Arithmetic Tasks (MAT) to properly stimulate different levels of stress. During the experiment, Photoplethysmogram (PPG) signals were recorded along with subjective feedback for validation of stress induction. The beat-to-beat interval series, estimated from sixty seconds long segments of PPG signals, were used to extract different features based on their reliability. In order to capture the temporal information in the ultra-short term segments of PPG, we introduced a new set of features which have the potential to quantify the temporal information at point-to-point level in the Poincare plot. We also used a Sequential Forward Floating Selection (SFFS) algorithm to mitigate the issues of irrelevancy and redundancy among features. We investigated two classifiers based on quadratic discriminant analysis (QDA) and Support Vector Machine (SVM). The results of the proposed method produced 94.33% accuracy with SVM for five-level identification of mental stress. Moreover, we validated the generalizability of the system by evaluating its performance on a dataset recorded with a different stressor (Stroop). In conclusion, we found that the proposed multilevel stress detection system in conjunction with new parameters of the Poincare plot has the potential to detect five different mental stress states using ultra-short term recordings of a low-cost PPG sensor.