Improved Affect Prediction Using Complexity Based Ultra-Short-Term Heart Rate Variability Features

Abstract

Heart rate variability (HRV) has been a useful tool for understanding human behavior. HRV features, derived from the inter-beat interval (RR) time series, reflect the autonomic nervous system processes of the body and have shown correlates with various mental processes. These processes include mental fatigue, workload, and anxiety, to name a few. Developing an understanding of these constructs in machines is key to improving human-computer interaction. However, HRV based emotion recognition is often limited to detection of negative (stress or anxiety) versus neutral emotional responses. Such systems when tested with subjects showing wider emotional responses may lead to errors. In addition to this, it is desirable for such emotion recognition systems to have high temporal resolution, thus allowing for almost real-time feedback and adaptive decision making. In this article, we explore the use of novel complexity-based feature set computed from so called ultra-short-term segments of 60 seconds. More specifically, we evaluate the potential of HRV features to distinguish stress vs. amusement vs. neutral vs. relaxation classes. Experiments using the WESAD database show that the proposed features extracted on ultra-short-term window of 60s and combined with benchmark features provide an overall improvement of 12.92 % balanced accuracy and 20 % F1-score over using only the benchmark features.