Use of Machine Learning to Identify Children with Autism and Their Motor Abnormalities.

The team filmed toddlers with a 3-D camera while they reached for a toy and dropped it in a bucket. The whole task took 30 seconds.

A computer learned the tiny arm-speed and angle patterns that only appeared in kids later diagnosed with autism. It then tried to spot those same patterns in new kids.

The program correctly picked out the toddlers with autism 96.7% of the time. It never needed to see the child’s face or hear a word.

Zhao et al. (2023) did the same trick with 7-second eye-movement clips and reached 87% accuracy. The shorter time and lower score show eyes work, but arm data is even clearer.

García-López et al. (2016) looked at the same reach motion yet found only some parts of the movement were off. Their paper describes the problem; Alessandro’s paper shows how to turn that problem into a quick screen.

da Silva et al. (2025) moved the idea out of the lab and into a nursery using eye-tracking. Accuracy dropped, but the trade-off is real-world ease. Together the studies chart a path from lab-perfect to clinic-ready.

You can’t strap motion-capture markers on every toddler, but the study proves motor signs are strong enough to act as a red flag. When you see clumsy, jerky reaching during play or table work, note it. Share the clip with your assessment team. Quick motor screens may soon sit beside M-CHAT and ADOS, giving you one more low-cost data point before the long wait for a full evaluation.