Major dataset reveals critical inconsistencies in how medical experts identify life-threatening brain events in vulnerable infants.
Every year, thousands of newborns experience seizures in neonatal intensive care units—silent brain storms that can cause permanent damage if not caught quickly. A comprehensive new dataset tracking 79 critically ill infants at Helsinki University Hospital reveals a troubling reality: even top medical experts frequently disagree on whether a seizure is actually happening, missing crucial opportunities for intervention.
The research, spanning over 6,000 hours of brain monitoring, found that experts identified seizures in vastly different numbers—ranging from 300 to 600 seizures in the same dataset. This dramatic variation isn't just an academic concern. When doctors can't reliably spot seizures on the squiggly lines of EEG readouts, vulnerable newborns may not receive time-sensitive treatments that could prevent lasting neurological damage.
The dataset has already been downloaded nearly 300,000 times by researchers worldwide, driving urgent development of AI systems that could standardize seizure detection. With automated tools showing promise in early trials, the medical community is racing to solve a problem that affects roughly 1-5 per 1,000 live births—ensuring no newborn's critical brain event goes unnoticed in the crucial first days of life.
Number of seizures identified by each expert shows significant disagreement
Most monitoring sessions lasted 60-90 minutes, with some extended cases
When doctors can't reliably spot seizures, vulnerable newborns may not receive time-sensitive treatments that could prevent lasting neurological damage.
Early AI models trained on this data show promise in reducing detection variability between hospitals.
Insurance coverage for AI-assisted monitoring is under review in several countries.
The research highlights broader challenges in medical AI development and expert agreement in critical care.
Share this story