Facial movements as biomarkers for autism: A Bayesian prevalence and machine-learning proof-of-concept study

Autism research has long sought reliable behavioural markers to support early identification and diagnosis. In recent years, researchers have increasingly recognised the potential of facial movements as behavioural markers for autism. To make progress, it is necessary to (1) identify which facial movement differences are most prevalent, (2) explore the utility of these movement differences in training machine learning classifiers; and (3) examine how classifier performance varies across subgroups historically underserved in autism screening and diagnosis. This proof-of-concept study utilized data from 25 autistic and 26 age-, gender-, and IQ-matched non-autistic adults who posed emotional facial expressions across two conditions (4,896 recordings). Bayesian analyses showed that the estimated prevalence of facial movement differences ranged from 41.05% to 83.16%. We constructed machine-learning classifiers trained on the eight most prevalent differences (70%+) to distinguish autistic and non-autistic participants. Our final, optimized K-Nearest Neighbours algorithm achieved 94.20% accuracy, with 94.64%/93.75% recall, and 93.81%/94.59% precision (autistic/non-autistic). Performance remained high across sexes and racial groups, with at least 93.75% accuracy, 95.63% recall, and 86.36% precision. Here, we demonstrated proof-of-concept that facial movement differences can effectively distinguish autistic and non-autistic people, suggesting that these movements may serve as reliable, objective behavioural markers for autism.