Automated vs Human Recognition of Emotional Facial Expressions of High-Functioning Children with Autism in a Diagnostic-Technological Context: Explorations via a Bottom-Up Approach

Early detection of autism spectrum conditions (ASC) is an important goal. Automated facial expression recognition is a promising approach and has implications for assistive and educational technologies, too. This study was an initial exploration of (1) the inter-rater reliability of human recognition of facial emotions of high functioning (HF) children with ASC; (2) the relationship between human and automated recognition of facial emotions; and (3) a ‘bottom-up’ approach on identifying ASC/typical development (TD) differences, from a screening serious game context. Thirteen HF, kindergarten-age children with ASC and 13 children with TD, matched along age and IQ, participated. Emotion recognition was administered on video-recordings from sessions of their playing with the serious game. Results showed lack of inter-rater reliability in human coding, confirming some advantages of machine coding. The simple bottom-up cross-sectional exploratory analysis did not reveal any ASC/TD difference. This is in contrast with our and others’ previous results, indicating such differences when aggregating emotion data from wider time-windows in machine-coded data-sets. This suggests that this second approach may be a more promising one to identify autism-specific emotion expression patterns.