Autism classification from public rendered skeleton videos: a participant-level comparative study

Public skeleton renderings enable reproducible research on video-based autism classification, but compression and small participant cohorts constrain both learning and evaluation. We investigate ChromaPose-MIL, a compact architecture combining soft chromatic decomposition, factorized spatiotemporal convolutions, temporal attention, and session-level multiple-instance pooling. The classifier is trained from random initialization on an Apple Silicon graphics processor using 160 children, equally divided between autism and typical development. Child-level five-fold cross-validation includes an inner validation partition and a fixed decision threshold. The primary model achieved an area under the receiver operating characteristic curve of 0.605 (conditional 95% interval 0.517–0.692), accuracy of 56.25%, sensitivity of 55.00%, and specificity of 57.50%. A static recording-layout control achieved an area of 0.532. Exploratory nested selection of classifiers on frozen visual features achieved 67.50% accuracy. A full-session neural pooling follow-up achieved 60.63% and did not improve that result. The study demonstrates local computational feasibility and documents how participant-level validation limits conclusions drawn from extensive video data. The observed discrimination and adaptive reuse of one cohort require independent replication before screening use can be considered.