Machine learning classification of autism spectrum disorder based on reciprocity in naturalistic social interactions

Abstract

Autism spectrum disorder is characterized by impaired social communication and interaction. As a neurodevelopmental disorder typically diagnosed during childhood, diagnosis in adulthood is preceded by a resource-heavy clinical assessment period. The ongoing developments in digital phenotyping give rise to novel opportunities within the screening and diagnostic process. Our aim was to quantify multiple non-verbal social interaction characteristics in autism and build diagnostic classification models independent of clinical ratings. We analyzed videos of naturalistic social interactions in a sample including 28 autistic and 60 non-autistic adults paired in dyads and engaging in two conversational tasks. We used existing open-source computer vision algorithms for objective annotation to extract information based on the synchrony of movement and facial expression. These were subsequently used as features in a support vector machine learning model to predict whether an individual was part of an autistic or non-autistic interaction dyad. The two prediction models based on reciprocal adaptation in facial movements, as well as individual amounts of head and body motion and facial expressiveness showed the highest precision (balanced accuracies: 79.5% and 68.8%, respectively), followed by models based on reciprocal coordination of head (balanced accuracy: 62.1%) and body (balanced accuracy: 56.7%) motion, as well as intrapersonal coordination processes (balanced accuracy: 44.2%). Combinations of these models did not increase overall predictive performance. Our work highlights the distinctive nature of non-verbal behavior in autism and its utility for digital phenotyping-based classification. Future research needs to both explore the performance of different prediction algorithms to reveal underlying mechanisms and interactions, as well as investigate the prospective generalizability and robustness of these algorithms in routine clinical care.

Fig. 1. Experimental setup.

Participants were seated across from each other and asked to conduct two conversational tasks. For additional setup info see Supplementary Material.

Fig. 2. Contribution of features in FACEsync model.

Cross-validation ratio of feature weights (A) and sign-based consistency (B) for the FACEsync model. The features depicted correspond to the person-specific adaptation of intensity of a participant to their dyadic counterpart in the respective facial action units (AU) for either hobbies or mealplanning task (min minimum, sd standard deviation, max maximum).

Fig. 3. Classification metrics for all base and stacking models.

BAC balanced accuracy, AUC area under the curve, PPV positive predictive value, NPV negative predictive value. Models are depicted in the order of lowest to highest performing BAC.