The Seizures No One Can See: How Well Do We Really Detect Them in Newborn Babies?

Two studies from 2025 — one measuring how accurate the brain monitor at the cot side is, one testing whether a computer can read brain waves as well as a specialist

Newborn babies can have seizures that are completely invisible — no twitching, no crying, nothing a nurse standing at the cot could see. Two studies published in 2025 asked how well these seizures are actually being detected. One found that the simplified brain monitor most units own misses roughly three in every ten affected babies. The other showed that a computer program can now read a baby's brain waves about as accurately as a human specialist.

Why an invisible seizure is such a difficult problem

The two studies described here — a large review of how accurate the bedside brain monitor is [1], and a study of a computer program trained to read brain waves [2] — both begin from the same difficulty. A seizure is a burst of disorganised electrical activity in the brain. In older children and adults it usually announces itself — shaking, stiffening, a loss of awareness. In newborn babies, and especially in babies who are already unwell, the link between the electrical event and any visible movement is often broken. The brain can be seizing while the baby lies perfectly still.

This matters because seizures in the first days of life are common in exactly the babies who are already at risk. They most often follow a period of reduced oxygen or blood flow around the time of birth — a condition doctors call hypoxic-ischaemic encephalopathy — or a stroke [3],[4]. Because the seizures are hard to see, the only way to know they are happening is to watch the brain's electrical activity directly, with an electroencephalogram, or EEG.

How families and clinicians used to face this

For a long time the answer was simply to watch the baby. Nurses and doctors looked for movements that might be seizures and treated what they saw. Then, about twenty years ago, studies that recorded brain activity continuously while staff watched at the bedside showed how much was being missed. The electrical seizure activity and the visible signs turned out to be only loosely connected, and a great deal was going unrecognised [5]. Later work following hundreds of monitored babies confirmed that many of them carried a heavy seizure burden and needed more than one medication to control it [6].

The obvious response — put a full EEG on every at-risk baby — ran into a practical wall. A full EEG, the kind that counts as the gold standard, uses at least nine electrodes on the scalp plus a video camera, and it has to be interpreted by a specialist trained in reading newborn brain traces. Most neonatal units cannot provide that around the clock. Some cannot provide it at all.

So the field adopted a compromise: amplitude-integrated EEG, usually shortened to aEEG. It uses just two to four electrodes and squeezes hours of brain activity into a single compressed line on a screen, designed so that a neonatal doctor or nurse — rather than a brain specialist — can read it at the bedside [7]. It spread quickly through neonatal units worldwide. What did not keep pace was proof of how accurate it is. Early studies raised doubts, and a review published in 2015 found that the answer depended enormously on details like whether the reader could also see the underlying raw brain-wave trace [8].

The first study: putting a number on the bedside monitor

The first of the two 2025 papers is a Cochrane review — the kind of study that gathers every piece of relevant research and analyses it together, using strict rules laid out before the search begins [1]. Its authors searched the medical literature without language restrictions, screened 764 studies, and found 16 that met their standard: 562 babies who had been monitored with the simplified aEEG and the full gold-standard EEG at the same time, so the two could be compared directly.

They asked two separate questions, and the distinction turns out to matter a great deal.

Does this baby have seizures at all? Pooling 13 studies covering 490 babies, aEEG correctly identified 71 out of every 100 babies who genuinely were having seizures, and correctly cleared 84 out of every 100 who were not. Turned around, that means about three in ten affected babies would be missed, and about one in six unaffected babies would be wrongly flagged.

How many seizures is this baby having? Here the researchers refused to give a single number, and explained why: seizures within one baby are not independent events, so averaging them across a group is statistically misleading. What they could report is the spread across studies — the proportion of individual seizures that aEEG caught ranged from none at all to 86 in every 100. The proportion of "seizures" flagged on aEEG that turned out to be false alarms ranged just as widely, from none to all of them.

The review also rated the strength of the evidence as low, and included two findings that should give pause. When the analysis was restricted to the three best-designed studies, accuracy went down, not up. And only two of the 16 studies had clinicians read the monitor live at the bedside, which is how it is really used; in those two, the detection rates were zero and 57 in every 100.

Reassuringly, the review also identified why things go wrong, and most of the reasons can be addressed. Seizures were missed when they were short, when they happened in a part of the brain far from where the few electrodes were placed, or when the reader was inexperienced. False alarms came from muscle movement, from patting the baby, from hiccups, and from electrodes that were not stuck on properly.

The second study: teaching a computer to read the trace

The second paper approaches the same problem from the other side [2]. If the real bottleneck is the shortage of specialists who can read a full EEG, then perhaps a computer could do the reading.

A research group working with recordings from Cork University Maternity Hospital in Ireland trained a type of artificial intelligence called a neural network on more than 50,000 hours of newborn brain-wave recordings. Crucially, two specialist neurophysiologists had gone through those recordings and marked every seizure, one electrode channel at a time — 12,402 separate seizure events across 77 babies. This is a far larger and far more detailed teaching set than anything used before.

The team then tested the trained program on two completely separate groups of babies it had never encountered: 51 from Cork and 79 from an openly shared research collection from Helsinki, Finland. Each of those recordings had already been read independently by three EEG experts.

The test they applied is an elegant one. Rather than asking "how often is the computer right?", they asked: if we quietly swap the computer in for one of the three human experts, does the group's level of agreement change? On both sets of babies, it did not. Swapping in the machine left agreement statistically unchanged. The authors describe this as the first thorough demonstration of expert-level automated seizure detection in newborns.

The program had other useful properties. It reliably caught every long seizure — those lasting more than five minutes — although short ones under 30 seconds accounted for most of its misses. Its estimate of a baby's total seizure time closely matched what the experts agreed on. And because it had been taught to read each electrode channel separately, it kept working even when half the channels were lost, which happens routinely when electrodes come loose on a moving baby.

What this means for families

If your baby is being monitored in a neonatal unit, none of this means the monitor at the cot side is useless. The review's own advice is measured: where full EEG is available, it should be used; where it is scarce, the simplified monitor is a sensible way of deciding which babies most need it; and where full EEG is not available at all, the simplified monitor is still better than watching alone.

What it does mean is that the team looking after your baby is working with a degree of genuine uncertainty, and that the honest version of their explanation includes it. A clean-looking monitor does not completely rule out seizures. A flagged event is not always a seizure. This is one reason clinical judgement and the whole picture of how a baby is behaving still matter alongside the screen.

It also helps explain a frustrating piece of the story. A trial called the Newborn Electrographic Seizure Trial (NEST), led from Monash University in Melbourne, Australia, and run in specialist newborn intensive care units, randomly assigned 212 babies to have monitor-detected seizures treated or not, and found no clear difference in survival free of severe disability at age two [9]. Its own investigators pointed out the difficulty: if the test defining who gets treated is imprecise, the trial cannot give a clean answer. Meanwhile, studies across multiple hospitals show that even when continuous monitoring is running, only a small minority of seizures are treated within an hour of starting [10] — and how much seizure activity a baby has is linked to how they do later [11].

What researchers are working on next

The immediate goal is to move automated reading from the research archive to the cot side. That has been attempted once already: the Algorithm for Neonatal Seizure Recognition (ANSeR) trial, run across eight neonatal centres in Ireland, the Netherlands, Sweden and the UK, tested an earlier detection program and did not find a clear improvement in how accurately clinicians identified seizures [12]. The program tested there dated from 2011 and has since been overtaken. What has also improved is the way such claims are checked — shared reference recordings and agreed statistical tests now let independent groups verify each other's results [13].

A 2026 review of artificial intelligence in newborn seizure detection concluded that most of the published work so far is still preliminary and that carefully conducted studies in real clinical settings are needed before these tools can be relied on [14]. Other teams are extending the same approach beyond seizures, training programs to judge how mature a baby's brain activity looks and to identify sleep states, with built-in checks that flag when the recording is too noisy to trust [15]. The honest position today is that a tool good enough to close this gap now exists and has been carefully validated on stored recordings — but the study that would prove it helps real babies in real units has not yet been done.

References

  1. Rakshasbhuvankar AA, Nagarajan L, Zhelev Z, Rao SC. Amplitude-integrated electroencephalography compared with conventional video-electroencephalography for detection of neonatal seizures. Cochrane Database of Systematic Reviews. 2025;8(8):CD013546. doi:10.1002/14651858.CD013546.pub2 ↩
  2. Hogan R, Mathieson SR, Luca A, Ventura S, Griffin S, Boylan GB, O'Toole JM. Scaling convolutional neural networks achieves expert level seizure detection in neonatal EEG. npj Digital Medicine. 2025;8(1):17. doi:10.1038/s41746-024-01416-x ↩
  3. Soul JS. Acute symptomatic seizures in term neonates: etiologies and treatments. Seminars in Fetal and Neonatal Medicine. 2018;23(3):183–190. doi:10.1016/j.siny.2018.02.002 ↩
  4. Pisani F, Spagnoli C, Falsaperla R, Nagarajan L, Ramantani G. Seizures in the neonate: a review of etiologies and outcomes. Seizure. 2021;85:48–56. doi:10.1016/j.seizure.2020.12.023 ↩
  5. Murray DM, Boylan GB, Ali I, Ryan CA, Murphy BP, Connolly S. Defining the gap between electrographic seizure burden, clinical expression and staff recognition of neonatal seizures. Archives of Disease in Childhood — Fetal and Neonatal Edition. 2008;93(3):F187–F191. doi:10.1136/adc.2005.086314 ↩
  6. Glass HC, Shellhaas RA, Wusthoff CJ, et al. Contemporary profile of seizures in neonates: a prospective cohort study. The Journal of Pediatrics. 2016;174:98–103.e1. doi:10.1016/j.jpeds.2016.03.035 ↩
  7. Hellström-Westas L. Amplitude-integrated electroencephalography for seizure detection in newborn infants. Seminars in Fetal and Neonatal Medicine. 2018;23(3):175–182. doi:10.1016/j.siny.2018.02.003 ↩
  8. Rakshasbhuvankar A, Paul S, Nagarajan L, Ghosh S, Rao S. Amplitude-integrated EEG for detection of neonatal seizures: a systematic review. Seizure. 2015;33:90–98. doi:10.1016/j.seizure.2015.09.014 ↩
  9. Hunt RW, Liley HG, Wagh D, et al. Effect of treatment of clinical seizures vs electrographic seizures in full-term and near-term neonates: a randomized clinical trial. JAMA Network Open. 2021;4(12):e2139604. doi:10.1001/jamanetworkopen.2021.39604 ↩
  10. Rennie JM, de Vries LS, Blennow M, et al. Characterisation of neonatal seizures and their treatment using continuous EEG monitoring: a multicentre experience. Archives of Disease in Childhood — Fetal and Neonatal Edition. 2019;104(5):F493–F501. doi:10.1136/archdischild-2018-315624 ↩
  11. Alharbi HM, Pinchefsky EF, Tran MA, et al. Seizure burden and neurologic outcomes after neonatal encephalopathy. Neurology. 2023;100(19):e1976–e1984. doi:10.1212/WNL.0000000000207202 ↩
  12. Pavel AM, Rennie JM, de Vries LS, et al. A machine-learning algorithm for neonatal seizure recognition: a multicentre, randomised, controlled trial. The Lancet Child & Adolescent Health. 2020;4(10):740–749. doi:10.1016/S2352-4642(20)30239-X30239-X) ↩
  13. Stevenson NJ, et al. Interobserver agreement for neonatal seizure detection using multichannel EEG. Annals of Clinical and Translational Neurology. 2015;2(11):1002–1011. doi:10.1002/acn3.249 ↩
  14. Lai NM, Yeo KT, Kong JY, et al. The use of artificial intelligence in neonatal seizure detection: an artificial intelligence-assisted systematic review. Journal of Paediatrics and Child Health. 2026. doi:10.1111/jpc.70513 ↩
  15. Hermans T, Dereymaeker A, Lemmens K, et al. Toward automated neonatal EEG analysis: multi-center validation of a reliable deep learning pipeline. Frontiers in Neuroscience. 2026;20:1750045. doi:10.3389/fnins.2026.1750045 ↩