Work

Autism Classification in Dyadic Interaction: An Interpretable Machine-Learning Benchmark Using Phase-Difference Motion Synchrony

Year: 2026
Type: article
Abstract: Interpersonal head-motion synchrony may differ in interactions involving adults with autism spectrum disorder (ASD), but its value for held-out dyadic classification remains uncertain. We reanalyzed 36 structured conversations involving 72 unique participants: 18 typically developing (TD) speaker-TD listener dyads (TD-TD) and 18 TD speaker-ASD listener dyads (TD-ASD). Four established phase-difference descriptors were extracted from 100 Hz forehead accelerometry. H1 predicted lower synchrony density and higher phase-difference standard deviation (SD) in TD-ASD dyads; Hedges g values were −1.20 and 1.01, respectively. For H2, seven classifier families evaluated with fixed and nested-cross-validation-tuned configurations showed no consistent nonlinear advantage (median fold-averaged AUCs, 0.731–0.815); equivalence between families was not established. For H3, density plus SD and density alone yielded leave-one-dyad-out AUCs of 0.793 and 0.790, vs. 0.765 for all four descriptors; these results do not demonstrate superiority, equivalence, or noninferiority. Synchrony events per speaker peak were less frequent in TD-ASD dyads (0.257 vs. 0.421), while chance-corrected coupling differences remained uncertain. These exploratory findings provide an interpretable, internally validated benchmark for prospective studies of dyadic behavioral decision support, not individual ASD diagnosis.
Language: English
Open Access status: diamond
OpenAlex ID: W7225892573
FWCI:
Cites: 50
Cited by:
Related to: 10