A Marked Point Process Approach for Continuous Valence Estimation Using Respiration Activity

Abstract

In this study, we present a method for continuously estimating emotional valence levels using a marked point process representation of features extracted from respiration amplitude signals. The amplitude of the breath, time of inhalation, and inhalation rate are used to label individuals breaths as potential pleasant or unpleasant valence events using an unsupervised k-means clustering algorithm. We generate two marked point processes consisting of both location and magnitude of inferred valence events corresponding to pleasant and unpleasant (high and low) changes in valence. A state-space model is then used to model high and low valence states based on the occurrence of events indicative of either state in each marked point process. The resulting high valence and low valence states are combined to yield a single estimate of valence level. The algorithm is tested on a dataset containing 23 participants viewing emotion-eliciting video clips. The estimation results for high and low periods, as identified by self-reported ratings, are then compared using a Wilcoxon signed rank test, showing that the method is capable of distinguishing high and low valence periods. The estimated valence level is generally able to capture the trends of the self-reported ratings for most subjects, but fails to fully capture rapid and drastic changes in valence. Continuously estimating valence levels can have applications in the monitoring of patients with mental disorders, such as clinical depression, or multimedia recommendation to identify trends and better develop control strategies to regulate emotions.

Keywords

• Affective computing
• biomedical signal processing
• emotion recognition
• respiration
• state estimation
• state-space modeling