Artificial intelligence tools used to analyze video recordings of young children playing with bubbles can identify subtle differences in the way children on the autism spectrum move their hands. Specifically, a study published in the journal Nature found that young children who were later diagnosed with autism moved their hands up and down more while flapping their wings than children who were not diagnosed. developmental science. These small changes in body movements may ultimately help clinicians recognize behavioral patterns in young children.
Repetitive hand flapping involves a child shaking or vibrating their hands or arms repeatedly over a period of time. This is a common behavior in young children, especially during moments of heightened excitement or stimulation. In older children and adults, frequent or vigorous hand flapping is often associated with autism spectrum disorders.
Both neurotypical children and children on the autism spectrum flap their hands, making it difficult to determine whether the behavior is a natural part of early emotional expression or an early sign of atypical development. The second year of life is a time of rapid developmental changes. Early diagnosis often relies on overlapping behavioral patterns that are difficult to quantify using standard observation alone.
Historically, researchers have studied these early behaviors by having clinical assistants watch hours of video recordings and manually record the beginning and end of a child’s movements. This manual annotation method tracks whether the behavior occurs and for how long. It cannot effectively capture the physical intensity, speed, or precise mechanical properties of the movement itself.
To find out whether the physical characteristic of hand flapping can help distinguish between typical and atypical behavior, the researchers examined video footage of children participating in structured play. Jan Stenham, a researcher in the Department of Physical Medicine and Rehabilitation at Johns Hopkins University School of Medicine, led the study with colleagues at the Kennedy Krieger Institute in Baltimore.
Researchers analyzed clinical assessments from a large historical study of early childhood development. The researchers focused on a final group of 28 infants between 13 and 16 months of age. Half of these children continued to show typical developmental milestones, but the other half later received a formal diagnosis of autism around age 3.
The researchers extracted a short three-minute segment of the video for each child. During this part of the clinical assessment, the examiner had the infant engage in standard bubble play activities. Bubbles routinely cause excitement and visual attention, and their activity is very likely to cause spontaneous hand flapping.
To assess specific mechanical properties of the infant’s movements, the team used a type of artificial intelligence called computer vision. Computer vision software trains computers to interpret visual information from digital images and videos. The researchers chose an algorithm known as AlphaPose, which automatically identifies and tracks anatomical landmarks in the human body.
When the software was applied to digital videos, digital coordinates were mapped to joints such as the infant’s shoulders, elbows, wrists, and fingertips. This created a moving digital stick figure that tracked the exact location of the child’s limbs pixel by pixel throughout the recording. This allowed the researchers to obtain objective data on the characteristics of children’s movements.
The team calculated two key measurements from the digital tracking data. They measured the amplitude, or the vertical distance the hand moves up and down during the flapping motion. They also calculated the frequency, which measures the rate of hand flapping per second.
The software adjusted the measurements based on each child’s physical proportions, as infants were sometimes standing close to and sometimes far from the camera lens. The researchers normalized the amplitude of hand flapping according to each infant’s torso height. This step ensures that the recorded data is not distorted by changes in body size or camera distance.
When the researchers analyzed each instance of hand flapping individually, they found measurable group distinctions. Infants who later received a diagnosis of autism showed higher movement amplitudes during individual flapping events compared to other children. Their hands traveled much longer vertical distances, reflecting more violent up-and-down physical movements.
Flap frequency did not follow this pattern. The researchers found no statistically significant differences between the two sets of infants in the speed of hand movements. Children in both groups vibrated their hands at approximately the same rate during each burst of excitement.
When the researchers averaged the data, rather than focusing on individual physical events, the differences disappeared completely. When the team combined all hand-flapping events for one child into one average score for that participant, the final numbers showed no statistically significant differences in either amplitude or frequency between the two groups.
The researchers suggested that this calculation effect may occur because infants’ behavior changes significantly from moment to moment. Children may flap their hands vigorously once they are excited, but then exhibit much smaller movements a minute later. Averaging all these instances erases the extreme highs and lows of the physical representation.
This finding highlights the importance of analyzing moment-by-moment behavior rather than relying on high-level summaries. Capturing these small dynamic expressions required the software to evaluate each short movement on its own merits, rather than blending them into a generalized behavioral profile.
The researchers noted several limitations in their study, which started with a small study sample of just 28 children. A larger group of infants will need to be evaluated to determine whether these elevated motor patterns emerge consistently in a broader population.
The study also focused on one behavior measured during a very specific three-minute play. Young children experience a variety of emotions and situations throughout the day. Long-term recordings across different environments could provide a more complete picture of how body movements fluctuate.
Additionally, half of the children in the comparison group had an older sibling diagnosed with autism. The chance that a sibling of an autistic child will eventually receive an autism diagnosis is approximately 20%. Even people who do not meet the full diagnostic criteria often exhibit early but subclinical features associated with autism, which can blur the statistical boundaries between the two groups.
Camera hardware also presents challenges and opportunities for future tracking efforts. The videos used in the study were collected over 20 years ago using standard equipment of the time. Modern smartphone cameras capture video at a much higher resolution. This means that future software applications will be able to track these joint coordinates with much higher accuracy.
Although automated video tracking holds promise for capturing subtle aspects of infants’ movements, the researchers cautioned that the technology is no substitute for clinical expertise. Direct observation, thorough developmental history, and in-person cognitive testing remain the standard methods. Instead, computerized tracking may ultimately provide an additional objective tool to assist clinicians during early childhood assessments.
The study, “Quantifying Repetitive Hand-Flapping Kinematics in Autistic and Non-Autistic Toddlers Using Video-Based Pose Estimation,” was authored by Jan Stenum, Elizabeth Ailer, Ryan T. Roemmich, Rebecca Landa, and Rachel Reetzke.
