Using AI facial expression analysis software to assess negative affect in autistic children

Assessing emotion regulation using traditional methods, such as questionnaires and observational tools, can be time-consuming, expensive, and prone to bias. AI-powered facial expression analytics holds promise in offering a more efficient and objective method for assessing emotion regulation – a cornerstone of mental wellbeing. However, the feasibility and reliability of using facial analytic software for treatment planning and patient monitoring remain unclear. This study addresses this knowledge gap by examining the clinical validity and accuracy of iMotions 10 AFFDEX SDK 5.1 in coding negative-expressed affect in autistic children. Fifty autistic children aged 7–12 years (44 males, 6 females) participated. Coding was performed on video footage of participants playing a video game in their homes for 5 min, followed by a 2-minute enjoyable task. Results indicated a strong correlation between the negative affect ratings of two human raters (r = .99, p < .001), and weak and non-significant correlations between the human raters and the AI algorithm (rs < .18, ps > .23). Analysis of signal data from AFFDEX SDK 5.1, when synchronised with participants’ video footage, suggested the software lacked sensitivity and misclassified anxious or frustrated facial grimacing as positive affect (smiling). These findings highlight the need for further validation of AI-driven technologies with specific clinical populations.