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

Background: Current screening tools for autism are ineffective and biased, leading to unnecessary referrals and delays in diagnoses, especially for women and people of colour. This has led autism community members and healthcare providers to call for more accurate, unbiased, and scalable screening tools. In recent years, researchers have increasingly recognised the potential of screening tools focused on facial movement differences between autistic and non-autistic people. To make progress, it is necessary to (1) determine which facial movement differences are most prevalent in autistic people, (2) train machine-learning classifiers on these features, and (3) optimise the classifier for subgroups underserved by screening and diagnosis. Method: This proof-of-concept study utilized data from 25 autistic and 26 age-, gender-, and IQ-matched non-autistic people who posed emotional facial expressions across two conditions, resulting in 4,896 recordings.Results: First, our Bayesian analyses showed that the estimated prevalence of facial movement differences ranged from 41.05% to 83.16%. Second, we constructed several machine-learning classifiers – trained on the eight most prevalent movement differences (70%+) to distinguish autistic and non-autistic participants. Third, we optimized our best-performing classifier through forward and backward feature selection techniques. Our final K-Nearest Neighbours algorithm achieved 91.87% accuracy, with 88.51% (autistic) and 95.89% (non-autistic) precision, and 96.25% (autistic) and 87.50% (non-autistic) recall. The classifier maintained consistently high performance, with at least 90.99% accuracy, 89.13% precision, and 87.42% recall, across sexes and racial groups. Conclusions: Here, we demonstrated proof-of-concept that facial movement differences can be used to effectively distinguish autistic and non-autistic people across sexes and racial groups. These results suggest movement-based classifiers could improve the accuracy and equity of autism screening. Future research should extend this approach to children to enable earlier identification and support.