Temporal dynamics of early child-clinician prosodic synchrony predict one year autism intervention outcomes using AI driven affective computing

The patient-therapist interpersonal dynamics is a cornerstone of psychotherapy, yet how it shapes clinical outcomes remains underexplored and difficult to quantify. This is also true in autism, where interpersonal interplay is recognized as an active element of intervention. Moreover, behavioral research is time-consuming and labor-intensive, limiting its translational applications. We studied 25 autistic preschoolers (17 therapists) across two naturalistic 60-minute sessions of developmental intervention at baseline and after three months (50 videos total). Clinical outcomes were assessed at baseline and one year into intervention. We developed a fully automated pipeline combining deep learning and affective computing to: (i) segment full-session audio recordings, (ii) model child-clinician acoustic synchrony using nonlinear metrics grounded in complex systems theory, and (iii) predict long-term response from early synchrony patterns. Changes in early synchrony dynamics predicted clinical response. Better outcomes were associated with synchrony patterns reflecting increased variability, predictability, and self-organization alongside prosodic features linked to emotional engagement. Our scalable, non-invasive system enables large-scale, objective measurement of therapy dynamics. In autism, our findings emphasize the importance of early interpersonal synchrony and emotional engagement as active drivers of developmental change. Our approach captures the full dynamics of entire therapy sessions, providing a richer, ecologically valid view of interpersonal synchrony.