Fine-Grained Analyses of Early Autism-related Social Behavior in Real-World Scenarios by Machine Learning

Autism Spectrum Disorder (ASD) is a condition that carries high costs for families and the healthcare system, requiring extensive management both in terms of diagnosis and treatment. The implementation of AI-based systems in clinical practice represents a possible supportive solution that can help clinicians by providing more systematic meth- ods to monitor child behavior. The main advantage over more traditional observational approaches is to offer quantitative and refined analysis solutions that can be ecological at the same time. The relevance of AI in clinical applications can have a role both in the challenge of early detection and in designing intervention programs better tai- lored to the specific functioning of children with ASD. The research project presented in this dissertation focused on developing AI-based systems for fine-grained analysis of autism-related social behaviors and their validation in concrete clinical environments. Specifically, in Chapter 2, our first study is presented, which targets on implementing a computational phenotyping system to address the need for new early markers of the condition. Through fine-grained analytics of facial dynamics in videos, we identified a set of features that distinguished young (6-12 months) infants with ASD (18 ASD, 15 non-ASD) during unconstrained at-home interactions. In Chapters 3 and 4, we introduce EYE-C, a Behavior Imaging model for robust analysis of eye contact episodes in eco- logical therapist-child interactions. The system was validated in the clinical setting for personalized early intervention. First, we investigated the influence of extracted features in categorizing spectrum heterogeneity across a sample of 62 preschool (<6 years) chil- dren with ASD. Further, we tested our metrics as predictors of early intensive treatment outcomes in a sub-sample of 18 subjects with ASD. The project aims to demonstrate the feasibility of effective computational systems that are robust to the high variability of unstructured interactions, with emphasis on the applicative value in real-world scenar- ios. Even though based on limited sample sizes, the work presented may offer interesting insights into the perspective of integrating AI into clinical practice. The research project was funded by an FBK scholarship and developed in a double in- ternship at ODFLab (University of Trento) and the FBK Data Science for Health (DSH) research unit.