Automated assessment of symptoms of anxiety in children from parent-child interaction videos

Abstract

Burgeoning economic crises, growing political instabilities, and the recent pandemic have caused mental health deterioration in many parts of the world. The most prevalent mental health affliction is anxiety disorder and it has been affecting a growing number of adults and children alike. In this thesis, we investigated the feasibility and approaches of applying machine learning for the detection of anxiety symptoms in 9-year-old children. These symptoms might be expressed more saliently through certain modes of communications and in specific interactional contexts. Hence, we experimented with models that were trained on unimodal and multimodal features extracted from video recordings of conflictual and cooperative interactions between nine-year-olds and their parents in a laboratory setting. Results suggest that anxiety symptoms manifest most noticeably during tense, conflictual interactions and are conveyed through the hand movements, facial expressions— particularly the mouth area—and word choice. Moreover, training with multimodal features demonstrated better performance compared to unimodal approaches. Although the resulting performance of the models was moderate, this study establishes the feasibility of detecting symptoms of anxiety using machine learning applied to multimodal dataset.