Tomorrow’s World: Artificial intelligence in neonatal brain monitoring and imaging (Session)

This was a strong session with three speakers who looked at AI from three angles. Geraldine Boylan covered what it can already do, Sampsa Vanhatalo covered what we should ask of it, and Petra Hüppi covered what imaging can tell us about mechanisms. Two of the three opened by saying “I’m not an AI expert”, which suited a session that cared more about clinical questions than about technology.

1. AI for neonatal seizure detection — Geraldine Boylan (Cork, Ireland)

The case. A term baby had a markedly discontinuous EEG from about 3 hours of age. Seizures followed in the middle of the night between Friday and Saturday. Full video-EEG was eventually set up and a neurologist confirmed seizures remotely. By the time the first drug was given, the baby was in status epilepticus and needed several antiseizure medications, including midazolam. On review, the first seizure had occurred at about 7.5 hours. It was low-amplitude and easy to dismiss as movement artefact. Could earlier detection have lowered the seizure burden and the number of drugs this baby needed?

Take-home messages

• Neonatal seizures are mostly invisible. Many are subclinical from the start, and electroclinical uncoupling after ASMs hides even more. EEG is the only reliable tool.

• The background warns you. In large cohorts, a severely abnormal or discontinuous early EEG identifies babies at higher seizure risk. Be extra vigilant in these babies.

• The first seizure is a red flag. Seizures can escalate quickly in frequency, spread and amplitude.

• Burden matters, and so does time to treatment. Higher total and peak seizure burden are linked to poorer cognitive and language outcomes at 18 months and to abnormal outcome in HIE. The number of minutes of seizures before first treatment matters, and earlier treatment reduces burden.

• There is only one real-time RCT so far. The ANSeR trial covered 8 European NICUs and used an algorithm from around 2014. Algorithm support increased the proportion of seizure hours detected from about 45% to about 66%.

• Algorithms have improved a lot. Open datasets such as the Helsinki dataset and much better models have given recent algorithms what she called expert-equivalent performance. They detect both seizure events and seizure duration, so they can estimate burden. They also work from a single channel up to a full montage, which helps with artefacts.

• Ask harder questions about metrics. AUC alone is not enough, because seizure data are heavily imbalanced and most of the recording is not seizure. Ask for sensitivity, specificity, PPV, NPV, false detections, agreement on seizure burden with experts, and inter-rater agreement. Testing must be on truly unseen data, with no patient overlap and data from many centres and devices. Two algorithms can have similar AUCs while one produces far more false alarms.

• Jurassic Park analogy. The raptors learned to open doors with lever handles. Show them a round knob and they’re stuck. An algorithm trained on one centre’s data behaves the same way.

• What comes next: regulatory approval, clinical translation, and treatment protocols ready to act on the first detected seizure, plus clinical trials. Technical accuracy alone does not establish clinical benefit.

2. AI-based monitoring of cortical recovery and maturation in neurocritical care — Sampsa Vanhatalo (Helsinki, Finland)

Take-home messages

• AI should not just emulate the expert. If it only reproduces what the EEG reader already says, it adds little. It should give clinicians more than that.

• Break down the bedside question. “How is the brain?” covers different time frames: what has happened, what is happening now, what will happen in the next hours, and how things evolve over hours to days and over weeks. Each needs a different approach and different underlying questions, such as suppression, encephalopathy grade, seizures, sleep state and maturity.

• Ground truth depends on the question. For seizures, background and sleep, models can learn from EEG experts. For maturation and prognosis, the ground truth must come from clinical information outside the EEG.

• Validation happens at two levels. The first is scientific validation for publication. The second is pragmatic: whether clinicians perceive the tool as valuable at the bedside. This matters most for adoption, but it is hard to measure and regulators don’t recognise it. One solution his group uses is a cloud platform where centres run the algorithms on their own data without sharing it.

• Background as a continuum. The brain is not “discontinuous” or “continuous” in discrete steps. Their model converts background into a continuous 0–100 score, which gives an objective multi-day recovery trend. Several studies link it to outcome.

• EEG-based “brain age”. Models trained on large preterm EEG datasets can estimate functional brain age and plot individual maturational trajectories. A preterm-age version was presented later at the meeting.

• An open question for clinicians: what error is acceptable, and what change in such a score is clinically meaningful? Engineers cannot answer this. The neonatal market is small, and getting from idea to certified product takes years and millions. Many algorithms are openly available for academic research, so use them and collaborate.

3. New horizons in neonatal brain imaging — Petra Hüppi (Geneva, Switzerland)

Take-home messages

• MRI still helps with counselling. In the case of “Hugo”, born at 26 weeks with severe IVH, MRI showed blood in the ventricles but intact parenchyma and metabolism. His parents chose to continue care.

• Faster scanning. MR fingerprinting, accelerated with machine learning, can produce several quantitative maps (T1, T2, synthetic contrasts, water fraction) in about 5 minutes, compared with roughly an hour of conventional sequences.

• Grey and white matter mature out of step in preterms. In healthy infants, T1 maturation of white matter and cortex track each other. In preterm infants, the cortex lags behind its white-matter connections.

• Low-field portable MRI (64 mT) makes imaging possible in low-resource settings, where most preterm babies are born. One example showed effects of maternal anaemia on infant brain growth, especially in the basal ganglia and corpus callosum.

• Socioeconomic factors weigh heavily. ML-based outcome prediction from MRI had only modest predictive value, and much of that depended on socioeconomic status. Preterm infants at term-equivalent age have reduced cortical folding, with fewer, shorter and shallower sulci. Clinical factors (birth-weight z-score, IUGR, multiple pregnancy, prolonged ventilation) explained about 3% of the folding variation. Maternal education explained 7–18%.

• Imaging can test interventions. In NICU music intervention studies, the music group had the most mature cortical folding. Interestingly, babies in an open bay did better than those in single rooms without parents present.

• Preterm birth affects specific networks. These include frontal, temporal, limbic and paralimbic connectivity, thalamic connectivity, the “rich club” hubs, and the salience, auditory and sensory networks. Variability in the BOLD signal between 33 and 40 weeks matches gene expression patterns in brain tissue at the same ages, which links imaging to biology. The mother’s singing and instrumental music engage different networks, so interventions can target specific networks.

• What’s coming next: glymphatic (brain waste-clearance) imaging and iron/myelin mapping at 7 T, and imaging of individual cortical layers.

Overall take-home

The algorithms are getting good. Seizure detection in particular is close to ready. What’s missing now is clinical: protocols that act on alarms, clear statements of which errors and changes matter, validation against real bedside questions, and trials showing patient benefit. Clinicians need to drive this.

Disclosure: This is an AI-assisted summary of the session, generated from a recording made by our team in attendance. We believe it is as accurate as our own notes would be, but please refer to the speakers’ original work before applying anything in clinical practice.