Multimodal, context-based dataset of children with Post Traumatic Stress Disorder

Abstract

The conventional method of diagnosing Post Traumatic Stress Disorder by a clinician has been subjective in nature by taking specific events/context in consideration. Developing AI-based solutions to these sensitive areas calls for adopting similar methodologies. Considering this, we propose a de-identified dataset of children subjects who are clinically diagnosed with/without PTSD in multiple contexts. This datset can help facilitate future research in this area. For each subject, in the dataset, the participant undergoes several sessions with clinicians and/or guardian that brings out various emotional response from the participant. We collect videos of these sessions and for each video, we extract several facial features that detach the identity information of the subjects. These include facial landmarks, head pose, action units (AU), and eye gaze. To evaluate this dataset, we propose a baseline approach to identifying PTSD using the encoded action unit (AU) intensities of the video frames as the features. We show that AU intensities intrinsically captures the expressiveness of the subject and can be leveraged in modeling PTSD solutions. The AU features are used to train a transformer for classification where we propose encoding the low-dimensional AU intensity vectors using a learnable Fourier representation. We show that this encoding, combined with a standard Multilayer Perceptron (MLP) mapping of AU intensities yields a superior result when compared to its individual counterparts. We apply the approach to various contexts of PTSD discussions (e.g., Clinician-child discussion) and our experiments show that using context is essential in classifying videos of children.