AI fetal heart rate monitoring has not yet been shown to improve outcomes for babies. The largest trial of computerised CTG interpretation, INFANT, randomised over 46,000 women and found poor neonatal outcomes in 0.7% of both arms. Newer deep-learning models look better on retrospective data, but none has proven it prevents harm in labour.
A labour ward at 3am. A midwife is watching a cardiotocograph, the paper or screen trace of a baby’s heart rate plotted against the mother’s contractions. The trace has a dip after each contraction. Is it reassuring, suspicious or pathological? Should she call the registrar, and should the registrar take the woman to theatre? Two experienced clinicians looking at the same trace may give different answers. Hindsight, in the form of a brain-injured baby and a lawsuit, will give a third.
If any task in medicine looks made for artificial intelligence, it is this one. CTG interpretation is pattern recognition on a continuous signal, performed by tired humans under time pressure, with catastrophic but rare errors. The data exist in enormous volume. And yet the history of automated CTG interpretation is one of the clearest cautionary tales in clinical AI.
The argument of this piece is that the problem was never only the reader. CTG itself is a weak test for the outcome everyone cares about, and software that reads a weak test faster or more consistently inherits its weakness. Any new AI system has to beat that ceiling, and it has to show it in randomised trials, not in retrospective accuracy curves.
Why is CTG interpretation so error-prone?
Continuous electronic fetal monitoring spread through high-income labour wards in the 1970s and 1980s on a reasonable theory: a baby short of oxygen will show it in its heart rate, and spotting that early allows delivery before damage is done.
The evidence for the theory turned out to be thinner than the practice. The Cochrane review by Alfirevic and colleagues, updated in 2017 and covering 13 trials with more than 37,000 women, compared continuous CTG with intermittent auscultation, listening at intervals. Continuous monitoring halved neonatal seizures (risk ratio 0.50). But it made no significant difference to cerebral palsy or perinatal death, and it raised caesarean sections (risk ratio 1.63) and instrumental births (risk ratio 1.15).
That pattern tells you what kind of test CTG is. It is sensitive to changes in heart rate, but most of those changes do not signal a baby in real trouble. The false positive rate is high, so clinicians who act on it perform many interventions that turn out to be unnecessary. And when real hypoxic injury does occur, the trace in hindsight often contained warning signs that were visible but not acted on, or were ambiguous among many similar traces that ended well.
On top of that, human interpretation is inconsistent. Studies have repeatedly shown high variability between observers, and even within the same observer on different occasions, in classifying the same trace. Guidelines have multiplied, and differ between countries and professional bodies. This is the environment in which software was supposed to help.
Did computerised CTG interpretation improve outcomes in trials?
It was tested properly, which is to the field’s credit, and it failed, which is the lesson.
The INFANT trial, published in The Lancet in 2017, randomised women in labour with continuous electronic fetal monitoring at UK and Irish maternity units. In one arm, clinicians had decision-support software that analysed the CTG in real time and raised alerts according to predefined rules. In the other, they had usual care. It was enormous: 46,042 women were analysed, 22,987 in the decision-support arm and 23,055 in the control arm.
The primary outcome, a composite of poor neonatal outcomes, occurred in 172 babies (0.7%) with decision support and 171 (0.7%) without, an adjusted risk ratio of 1.01. At two years, there were no differences in development. A later economic analysis alongside the trial found no significant differences in cost either.
It is worth dwelling on why. The software did roughly what it was designed to do: apply rules consistently and alert clinicians. But the rules encoded the same indicators humans already used, and the alerts arrived into a system where decisions about escalation, staffing, theatre availability and communication were the real bottlenecks. A better reader of a weak signal, dropped into a busy ward, did not change what happened to babies.
Is deep learning better than the older CTG software?
The new generation of AI differs in an important way. Instead of encoding guideline rules, deep-learning models learn directly from large archives of traces linked to outcomes, and can in principle find patterns no guideline describes. Researchers have also combined the heart rate signal with clinical risk factors, placental biomarkers and labour progress.
The retrospective results are more honest than the marketing sometimes suggests. A 2023 study in the journal Bioengineering trained a deep-learning model on the first 20 minutes of fetal heart rate recordings from 51,449 term births. At a specificity of 95%, meaning it wrongly flagged only 5% of healthy births, it detected about 20% of adverse outcomes. That is a real signal extracted from very early in labour. It also means four in five of the babies who went on to have problems were not flagged, even before anyone asks how clinicians would respond to the alerts.
Expert reviews in obstetric journals in recent years have argued for a shift from subjective CTG reading to computerised, quantifiable assessment, and for large shared databases with common evaluation standards before clinical rollout. That is the right direction. But it is also an admission that the field is still at the stage of building the evidence infrastructure, not of deploying proven tools.
There is a quieter point in the research too. Qualitative studies of midwives and obstetricians show that clinicians want AI-CTG systems to be accurate, transparent and independently evaluated before they will rely on them. That is not technophobia. It is the rational response of a profession that has already seen one generation of CTG software fail its trial.
A worked example: what an alert rate means on a real ward
Hypothetical, with round numbers for illustration. A unit delivers 5,000 babies a year, most with continuous monitoring. Suppose serious hypoxic injury affects about 1 in 1,000 births, so five babies a year.
An AI model that flags 5% of labours as high risk sends 250 alerts a year, roughly five a week. If it catches 20% of the babies who will be injured, it flags one of the five. The other 249 alerts concern babies who would have been fine.
Now the operational questions start. Does each alert trigger an obstetric review? Does the review lead to earlier caesarean in some of those 250 women, with the surgical and future-pregnancy risks that carries? Is the one true positive distinguishable from the rest at the moment the alert fires? And what happens to staff attention on the five-alerts-a-week ward when four of five babies the model misses show no warning at all?
Push the threshold down to catch more of the injured babies and the alert count rises steeply. Push it up and the model catches even fewer. There is no free setting. This is why the only question that matters is whether a system, together with the ward’s response to it, changes outcomes in a trial. INFANT asked exactly that question of the previous generation of software.
What is the strongest case for AI in fetal monitoring?
The optimistic case is not foolish, and it deserves a hearing.
First, INFANT tested rule-based software built on existing guidelines. Learning systems that combine the trace with clinical context are genuinely different, and one negative trial of an older approach does not settle the question for all future ones.
Second, consistency has value even without better accuracy. Inquiries into maternity failures in the UK and elsewhere have repeatedly identified CTG misinterpretation and failures to escalate as contributory factors. A system that flags a deteriorating trace to a senior clinician, and records that it did, could help with escalation and accountability even if it adds little diagnostic power.
Third, AI may add most in settings where there is currently little monitoring at all. In many low-resource labour wards, a single midwife may care for many women with intermittent listening. A low-cost monitor with automated interpretation that simply ensures heart rate is checked regularly could matter more than any refinement in a well-staffed unit. As we argued in our piece on a simple maternal health intervention with an outsized effect, the largest gains in maternal and newborn care often come from making basic care reliable, not from sophisticated additions.
These are reasonable arguments. None of them is yet evidence that babies do better. They are hypotheses to test, in the settings where they are claimed to work, with outcomes that matter.
What this means for maternity units, clinicians and vendors
For maternity units, the lesson of INFANT is to be sceptical of any product that claims to reduce brain injury or stillbirth on the strength of retrospective accuracy. Ask for randomised or at least well-controlled prospective outcome data. Ask what the alert rate will be on your ward, and who will respond to each alert. Treat escalation, staffing and theatre access as part of the intervention, because they are.
For obstetricians and midwives, AI outputs should be one input to a clinical picture that includes the woman’s history, labour progress and what she is telling you. Recording when you disagree with a system, and why, is valuable both for safety and for learning whether the system helps.
For vendors and researchers, report sensitivity at clinically realistic false positive rates, not only an AUC. Publish alert burdens. Design trials with outcomes that matter, including caesarean rates as a harm, and run them in the kinds of units that will buy the product. The general principle, that diagnostic AI is most convincing on narrow, well-defined questions, applies with force here; we explored it in our analysis of where AI diagnostic evidence is strongest.
For regulators and commissioners, CTG AI should be evaluated as a system intervention in labour care, not as a standalone device with a sensitivity figure. The question is whether outcomes change, not whether a curve looks good.
For women in labour, the practical point is that monitoring is a tool used by your care team, not a verdict. If you have questions about how you or your baby are being monitored, ask your midwife or obstetrician.
Frequently asked questions
Does AI improve fetal heart rate monitoring in labour?
Not yet proven. The INFANT trial of computerised CTG decision support, published in The Lancet in 2017 with more than 46,000 women, found poor neonatal outcomes in 0.7% of babies in both arms. Newer deep-learning systems show some promise in retrospective studies, but none has shown in a randomised trial that it reduces brain injury or death.
Why is CTG so hard to interpret?
CTG shows many heart rate changes that do not reflect real oxygen shortage, so false alarms are common, and clinicians disagree about how to classify the same trace. A Cochrane review found continuous CTG halved neonatal seizures but did not reduce cerebral palsy or perinatal death, while increasing caesarean births. Any software reading CTG inherits these limits.
How accurate are deep-learning CTG models?
Accuracy depends on the false alarm rate chosen. One 2023 study of 51,449 term births found a deep-learning model using the first 20 minutes of heart rate data detected about 20% of adverse outcomes while flagging only 5% of healthy births. That is a meaningful signal but far from a reliable screening tool on its own.
Could AI fetal monitoring help in low-resource settings?
Possibly, because many labour wards there have little monitoring and few staff. Automated tools that help ensure regular heart rate checks and prompt referral could matter more there than in well-staffed hospitals. This remains to be shown in trials, and it depends on having staff and facilities to act on alerts, including access to timely caesarean delivery.
Should I ask for AI monitoring during my labour?
AI-based CTG interpretation is not a standard of care and its benefit is unproven. What matters most is attentive monitoring by trained staff and a unit able to act quickly when concerns arise. Discuss any questions about monitoring with your midwife or obstetrician, who can explain what your unit uses and why.
Automated CTG is the rare corner of clinical AI where we already have a large, well-run trial, and it was negative. The next generation may do better. It should be held to the same test, because the price of over-promising here is paid in unnecessary surgery on one side and false reassurance on the other.
If you work on a labour ward, would you trust an AI alert enough to change what you do at 3am, and what would it need to show you first?
