AI Shows Early Promise in Detecting Infantile Spasms

LOS ANGELES — Artificial intelligence (AI) analysis of caregiver-recorded videos has the potential to diagnose infantile epileptic spasm syndrome, according to a new study.

Infants with the condition can have poor outcomes with even small delays in diagnosis and ensuing treatment, potentially leading to intellectual disability, autism, and worse epilepsy. “It’s super important to start the treatment early, but oftentimes, these symptoms are just misrecognized by primary care or ER physicians. It takes a long time to diagnose,” said Gadi Miron, MD, who presented the study on December 7 at the American Epilepsy Society (AES) 78th Annual Meeting 2024.

What Is This? What Should I Do?

Parents who observe unusual behavior often seek advice from friends and family members and receive false reassurance that such behavior isn’t unusual. Even physicians may contribute if they are unaware of infantile spasms, which is a rare disorder. “And then again, they get false reassurance, and because of that false reassurance, you get a diagnostic delay,” said Shaun Hussain, MD, who was asked to comment on the study.

The timing and frequency of infantile spasms create challenges for diagnosis. They only last about 1 second, and they tend to cluster in the morning. By the time a caregiver brings an infant to a healthcare provider, they may have trouble describing the behavior. “Parents are struggling to describe what they saw, and it often just does not resonate, or doesn’t make the healthcare provider think about infantile spasms,” said Hussain.

The idea to employ AI came from looking at videos of infants on YouTube and the realization that many patients upload them in an effort to seek advice. “So many parents upload these videos and ask in the comments, ‘What is this? What should I do? Can somebody help me?’ said Miron, who is a neurologist and researcher at Charité — Universitätsmedizin Berlin, Berlin, Germany.

AI and Video Can Aid Diagnosis

The researchers built a model that they trained to recognize epileptic spasms using openly available YouTube videos, including 141 infants, 991 recorded seizures, and 597 non-seizure video segments, along with a non-seizure cohort of 127 infants with an accompanying 1385 video segments.

Each video segment was reviewed by two specialists, and they had to agree for it to be counted as an epileptic spasm.

The model detected epileptic seizures with an area under the curve (AUC) of 0.96. It had a sensitivity of 82%, specificity of 90%, and accuracy of 85% when applied to the training set.

The researchers then tested it against three validation sets. In the first, a smartphone-based set retrieved from TikTok of 26 infants with 70 epileptic spasms and 31 non-seizure 5-second video segments, the model had an AUC of 0.98, a sensitivity of 89%, a specificity of 100%, and an accuracy of 92%.

A second smartphone-based set of 67 infants, drawn from YouTube, showed a false detection rate of 0.75% (5 detections out of 666 video segments). A third dataset collected from in-hospital EEG monitoring of 21 infants without seizures revealed a false-positive rate of 3.4% (365 of 10,860 video segments).

The group is now developing an app that will allow parents to upload videos that can be analyzed using the model. Physicians can then view the video and determine if there is suspicion of a seizure.

Miron also believes that this approach could find use in other types of seizures and populations, including older children and adults. “We have actually built some models for detection of seizures for videos in adults as well. Looking more towards the future, I’m sure AI will be used to analyze videos of other neurological disorders with motor symptoms [such as] movement disorders and gait,” he said.

Encouraging Early Research

Hussain, who is a professor of pediatrics at UCLA Health, lauded the work generally but emphasized that it is still in the early stage. “Their comparison was a relatively easy one. They’re just comparing normal versus infantile spasms, and they’re looking at the seizure versus normal behavior. Usually, the distinction is much harder in that there are kids who are having behaviors that are maybe other types of seizures, which is much harder to distinguish from infantile spasms, in contrast to just normal behaviors. The other mimic of infantile spasms is things like infant heartburn. Those kids will often have some posturing, and they often will be in pain. They might cry. That’s something that infantile spasms will often generate, so that’s why there’s a lot of confusion between those two,” said Hussain.

He noted that there have been efforts to raise awareness of infantile spasms among physicians and the general public, but that hasn’t reduced the increased detection.

Another Resource

In fact, parents with suspicions often go to social media sites like YouTube and a Facebook group dedicated to infantile spasms. “You can Google infantile spasms, and you’ll see examples of weird behaviors, and then you’ll look in the comments, and you’ll see this commenter said, ‘These could be infantile spasms. You should go to a children’s hospital. Don’t leave until you get an EEG to make sure that these are not seizures. There’s all kinds of great advice there, and it really shouldn’t be the situation where to get the best care, you need to go on YouTube,’” said Hussain.

Miron and Hussain had no relevant financial disclosures.