鶹ýӰ

AI Screening of Home Videos Proves Accurate for Infantile Seizure Detection

— But artificial intelligence model needs further validation before release

MedpageToday

LOS ANGELES -- An artificial intelligence (AI) model accurately detected and ruled out infantile epileptic spasm syndrome in videos captured by family, a study showed.

The model had an area under the receiver-operating-curve (AUC) of 0.96, with 82% sensitivity and 90% specificity, reported Gadi Miron, MD, of Charité-Universitätsmedizin Berlin in Germany, at the American Epilepsy Society annual meeting. The results were simultaneously published in a .

Validation on external datasets from smartphone videos yielded a similar AUC of 0.98 with a false positive rate (FPR) of 0.75%, as did comparison to gold-standard video-EEG (AUC 0.98, FPR 3.4%).

The tool has great promise if further developed into a simple, rapid screening tool for families or physicians to use, commented Deborah Holder, MD, of Cedars-Sinai Guerin Children's in Los Angeles.

"When babies have this diagnosis, we consider it a neurologic emergency, so anything we can do to make a fast diagnosis and get these babies to medical treatment is really important," she told MedPage Today.

Delayed diagnosis is common due to misrecognition of symptoms and can lead to long-term cognitive problems, Miron agreed.

The idea came from a trove of YouTube videos of potential epileptic spasms that parents had posted asking, "'What is this? Who can help me?'" Miron said. "There's a very active community of these parents trying to help each other out, and the clinical need is just basically coming from that."

Pediatricians and primary care physicians are often presented with these videos recorded by parents, since the events can come in clusters, allowing parents time to catch them on video. "They don't always happen when you go to the doctor's clinic, and then the parents need to describe and it's much harder to understand and to recognize what's happening," Miron noted.

If the doctor has a relationship with a neurologist or epilepsy specialist, the videos typically then get sent for a quick screen, noted Holder, who says she receives such videos every day. Otherwise, families have to wait for a clinic visit, which adds to the delay.

"There are not a lot of us around to look at them individually," she said. "There are whole areas of the country where there's just resources that aren't available. So to be able to help the primary care docs have better access to the expertise, and if there's computer programs that can be trained to look at these videos, it's very cool and very exciting. We just want to make sure that it's in the hands of the people who know how to use it and get the patients where they need to be."

Miron agreed that deployment will be key.

"This is still in the research stage, but I think it shows a lot of promise, because we have addressed some of the issues that are important in order for it to be clinically transferable," Miron said. "Of course, in order to be clinically applicable, it needs to be validated further prospectively in additional datasets. It needs to be matched with some type of tool that the parents can use."

His group built an app and a doctor's platform and are testing them in the clinic and emergency department in a prospective single-center study, with hopes to expand to other centers as well, he said.

His group trained the AI model on a YouTube video from 141 children with 991 seizures, found by searching for "epileptic spasms," "infantile spasms," and "West syndrome," compared with video from 127 children without seizures.

These were validated against two cohorts sourced from smartphone videos: 26 infants with seizures (70 seizure and 31 non-seizure 5-second video segments) and 67 infants without seizures. Validation against in-hospital video-EEG monitoring, included 21 infants without seizures to look for a false alarm rate.

"Videos demonstrate high heterogeneity in resolution, bit rate, brightness, and sharpness," Miron's group noted. "Our model performs well on multiple datasets, exhibiting robustness across camera sources and technical heterogeneity. Future studies should focus on validation, translation, and clinical application."

Disclosures

The study was partially funded by a grant from the Berlin Institute of Health.

Gadi and Holder disclosed no relationships with industry.

Co-authors disclosed patent applications related to neurological conditions outside the study, and relationships with Angelini, Bial, Desitin, Eisai, Jazz Pharma, Neuraxpharm, Nutricia, and UCB.

Primary Source

American Epilepsy Society

Miron G, et al "Detection of epileptic spasms using foundational AI and smartphone videos: A novel diagnostic approach for a rare neurological disorder" AES 2024; Abstract 1.179.