To rank or to classify? Annotating stress for reliable PTSD profiling

Abstract

In this paper we profile the stress responses of patients diagnosed with post-traumatic stress disorder (PTSD) to individual events in the game-based PTSD stress inoculation and exposure virtual environment StartleMart. Thirteen veterans suffering from PTSD play the game while we record their skin conductance. Game logs are used to identify individual events, and continuous decomposition analysis is applied to the skin conductance signals to derive event-related stress responses. The extracted skin conductance features from this analysis are used to profile each individual player in terms of stress response. We observe a large degree of variation across the 13 veterans which further validates the idiosyncratic nature of PTSD physiological manifestations. Further to game data and skin conductance signals we ask PTSD patients to indicate the most stressful event experienced (class-based annotation) and also compare the stress level of all events in a pairwise preference manner (rankbased annotation). We compare the two annotation stress schemes by correlating the self-reports to individual event-based stress manifestations. The self-reports collected through class-based annotation exhibit no correlation to physiological responses, whereas, the pairwise preferences yield significant correlations to all skin conductance features extracted via continuous decomposition analysis. The core findings of the paper suggest that reporting of stress preferences across events yields more reliable data that capture aspects of the stress experienced and that features extracted from skin conductance via continuous decomposition analysis offer appropriate predictors of stress manifestation across PTSD patients.

The Ordinal Nature of Emotions: An Emerging Approach

“…The breadth of applications expand to video-based [57], [76], [77], image-based [35], [37], [37], speech-based [106], music-based [70], [110] or physiology-based [111] emotion recognition for health, educational or entertaining [78] purposes. The next section covers a few successful applications directly showcasing the benefits of ordinal annotation and processing for affect modeling.…”

“…Figure 7 showcases how much closer a preference learned model can reach a hypothesized (artificial) ground truth, compared to a classification model. Importantly for the thesis of this paper, Holmgaard et al [111] compare different types of stress annotation with the aim of finding the best possible approximation to the underlying ground truth. In particular they compare, in a first-order manner, annotations indicating the most stressful event in a game (class-based annotation) versus a rankbased approach by which subjects compare stress across game events.…”

Towards general models of player affect

“…This final feature value represents the relative change of the statistic with respect to a particular participant's baseline. Beyond these standard statistical features, and inspired by the study of Holmgård et al [27], we applied a continuous decomposition analysis (CDA) [28] to the EDA signal using Ledalab. The outcome of CDA is the decomposition of the EDA in its phasic and tonic components.…”

Toward automatic detection of acute stress: Relevant nonverbal behaviors and impact of personality traits

“…Several systems have been developed for the detection of stressful states. These systems and their evaluation make use of the analysis of physiological signals, including blood volume pressure [5], heart rate variability (HRV) [6], skin conductance [7], and cortisol saliva samples [8]. However, most of these systems require invasive sensors that may themselves induce stress in participants.…”