When a toddler bursts a bubble, the motion of their hands may hold clues that only artificial intelligence can uncover. A new study in Developmental Science shows that a computer‑vision system can pick up subtle differences in how toddlers move their hands while playing with bubbles.

The research, carried out by scientists at Johns Hopkins University and the Kennedy Krieger Institute, examined 28 children between 13 and 16 months old. Half of the children were later given a formal autism spectrum disorder (ASD) diagnosis around age three; the other half reached typical developmental milestones.

Each child was recorded for a three‑minute bubble‑play session—an activity that naturally elicits excitement and spontaneous hand flapping, a common behavior in toddlers that can also signal ASD in older children. Using the AlphaPose algorithm, the team automatically tracked key body joints—shoulders, elbows, wrists, and fingertips—frame by frame, producing a digital stick‑figure of each child’s movements.

From the joint data the researchers calculated two metrics for each hand‑flapping event: amplitude, the vertical distance the hand travels, and frequency, the number of flaps per second. Because toddlers sat at varying distances from the camera, amplitude was normalized to each child’s torso height to reduce size‑related bias.

When the investigators examined individual flapping events, they found that children who later received an ASD diagnosis exhibited a higher amplitude. Their hands moved through a larger vertical distance during each flapping burst, while the frequency of flaps did not differ significantly; both groups flapped at roughly the same speed.

A key insight emerged when the data were aggregated. Averaging the amplitude or frequency across all flapping events for a single child erased the difference between the groups. The researchers explained that toddlers’ hand movements vary widely from moment to moment—one burst may be vigorous, another more modest—and that averaging smooths out these extremes, masking the distinctive pattern that appears when events are examined individually.

The authors emphasize that analyzing moment‑to‑moment behavior is essential for capturing the nuanced expressions that may signal atypical development. Manual annotation of video, the traditional method for studying early childhood behavior, records only whether a behavior occurs and its duration; it does not capture the intensity or precise mechanical qualities of the movement.

Several limitations were acknowledged. The sample size of 28 toddlers is small, and the study examined only one behavior in a single, highly structured activity. Longer recordings across varied settings could provide a more comprehensive view of hand‑flapping patterns. Additionally, half of the children in the comparison group had older siblings with an ASD diagnosis, which may introduce subclinical traits that blur group boundaries.

The video footage was collected more than twenty years ago with standard‑definition cameras. Modern smartphone cameras offer higher resolution, which could improve the precision of joint tracking in future studies.

While the research demonstrates the potential of AI‑driven video analysis to quantify subtle motor behaviors, the authors caution that the technology does not replace clinical expertise. Direct observation, developmental history, and cognitive testing remain the gold standard for diagnosing ASD. Instead, automated tracking could serve as an objective adjunct to assist clinicians during early evaluations.

The study, titled Quantifying Repetitive Hand Flapping Kinematics in Autistic and Non‑Autistic Toddlers Using Video‑Based Pose Estimation, was authored by Jan Stenum, Elizabeth Eiler, Ryan T. Roemmich, Rebecca Landa, and Rachel Reetzke.

In summary, AI‑based pose estimation can detect higher hand‑flapping amplitude in toddlers who later receive an ASD diagnosis, but the effect disappears when data are averaged. The findings underscore the importance of analyzing individual movement events and suggest that further research with larger samples and diverse settings is needed before such tools could be integrated into routine clinical practice.