Perinatal stroke — a stroke occurring before or around the time of an infant’s birth that can lead to lifelong disability — affects around 2 in 1,000 children, making it the most common motor disorder in childhood. Early intervention has the potential to improve outcomes, but it requires early detection, and that’s easier said than done. The symptoms tend to be nonspecific, and one routine screening method — General Movement Assessment (GMA), which relies on recognizing movements characteristic of a stroke — requires extensive training.
Fortunately, researchers at Newcastle University and the Georgia Institute of Technology believe they’ve made progress toward an automated, low-cost diagnostic solution that leverages a combination of wearables and artificial intelligence (AI). Their work, which they describe in a newly published preprint paper (“Towards Reliable, Automated General Movement Assessment for Perinatal Stroke Screening in Infants Using Wearable Accelerometers“), involves outfitting newborns with body-worn sensors and applying GMA algorithms to the data collected.
In small, preliminary tests involving 34 infants — 13 of which had abnormal movements — the researchers’ system identified likely cases of a perinatal stroke over 75 percent of the time.
“[It’s an] encouraging [advance] towards our ultimate goal of an automated PS screening system that can be used population-wide,” the paper’s authors wrote. “Our goal is to develop methods that enable population-wide, minimal-effort yet accurate and objective assessments of every newborn that will lead to early detection of potential motor abnormalities. Such an automated screening procedure will not lead to fewer cases of PS but to earlier detection.”
In the course of the study, the researchers attached lightweight, brightly colored cotton straps containing accelerometers to the infants’ ankles and wrists and recorded data during 10-minute trials with each newborn at monthly intervals. (They also captured video footage of the trials, which was reviewed separately by trained professionals.) Accelerometer readings from 161 total validated tests (between 24,000 and 60,000 samples) were fed through a machine learning-based pipeline (the Discriminative Pattern Discovery) to classify signs of abnormal movements that might indicate a stroke.
In the validation phase, their method achieved an accuracy of 80 percent, consistently outperforming prior approaches. Importantly, they note, the accuracy was higher than that required by trained human annotators in GM exams.
“[Our] developed Discriminative Pattern Discovery (DPD) method automatically detects relevant patterns and bootstraps effective classification models based on these,” the researchers wrote. “Through a rigorous evaluation of our method in a cohort of infants, who either had been diagnosed with Perinatal Stroke or were typically developing, we have laid the foundation for a screening tool that can potentially be used … with minimal effort to enable early-stage recognition of abnormal movements in young infants. Our method is straightforward to apply, inexpensive, and reliable with regards to the accuracy of analysis results.”