Early labour CTG model separates fetal risk with data already in the room
A new prognostic model built for the opening stretch of labour uses routinely available clinical variables and the first 20 minutes of cardiotocography to estimate the risk of severe fetal compromise at birth. The tool, called Fit4Labour, was developed from a large retrospective maternity dataset and then tested across two external validation cohorts with different populations and care environments. Its aim is straightforward and clinically urgent: replace delayed interpretation and subjective uncertainty with an objective estimate of fetal risk while labour is still unfolding.
The model combines maternal and intrapartum factors that obstetric teams already collect: maternal age, gestational age, nulliparity, pre-eclampsia, diabetes in pregnancy, thick meconium, and the highest maternal temperature recorded in labour. It then adds quantitative features extracted from the first 20 minutes of CTG tracing, including fetal heart rate baseline, short-term variability, decelerative capacity, number of accelerations, and number of prolonged decelerations. In practice, that creates an early warning system built on signals that reflect placental function, fetal autonomic reserve, and evolving stress during labour.
A single prognostic estimate at labour onset can change the tempo of decision-making.
Large multicentre datasets gave the model room to fail and room to prove itself
The development and validation work drew on 147,350 births with at least 20 minutes of CTG data. Oxford contributed 91,378 births spanning 1993 to 2024, with one segment used for training and testing and a later segment reserved for internal validation. Cambridge added 14,709 births from 2016 to 2022 for external validation, and Stockholm-Gotland contributed 41,263 births from 2018 to 2020 for independent external validation. Singleton births at 36 weeks’ gestation or later were included if they underwent intrapartum CTG monitoring, whether labour began spontaneously or was induced.
The outcome framework separated births into three mutually exclusive categories: severe compromise, intermediate compromise, and healthy. For clinical use, the model generated a Fit4Labour risk score representing the estimated probability of severe compromise and then divided women into four risk strata, from the highest-risk group above the 95th centile down to the lowest-risk group at or below the 25th centile. Performance was judged using partial receiver operating characteristic curves and compared against a clinical benchmark based on sensitivity at a 15% screen-positive rate.
Those design choices matter because intrapartum prediction tools often look impressive only inside the dataset used to build them. Here, the model had to travel across centres with different maternity demographics and differing clinical practice patterns. It did.
Severe compromise concentrated in the highest-risk group across validation cohorts
In every validation dataset, the highest-risk group carried the greatest burden of severe compromise, while the lowest-risk group remained at the bottom of the distribution. In Oxford’s 2020 to 2024 internal validation cohort, severe compromise occurred in 0.9% of the highest-risk group and 0.1% of the lowest-risk group. Cambridge showed 0.3% versus 0.1%, and Stockholm-Gotland showed 0.9% versus 0.2% in the same comparison.
Those absolute percentages are small, but they track in the direction that matters for bedside triage: the model enriched risk where it was supposed to and left the lowest strata with a substantially lower event rate. CTG alone often becomes noisy because the tracing is influenced by fetal sleep cycles, maternal medication, contractions, and observer interpretation. By integrating objective tracing features with fixed obstetric risk factors, Fit4Labour aims to capture a broader physiological picture of fetal reserve and decompensation.
The data also point to a model that can be interpreted in real time rather than hidden inside a black box. Quantitative CTG features can be measured during the first 20 minutes of monitoring, which means the score can, in principle, be generated early enough to affect surveillance intensity, escalation thresholds, or the decision to seek senior review.
Sensitivity did not beat the clinical benchmark, but performance remained stable across sites
At a fixed 15% screen-positive rate, 20-minute Fit4Labour achieved a sensitivity of 33.4% with a 95% confidence interval of 28.31 to 38.56. The clinical benchmark sensitivity was 39.9%, with a 95% confidence interval of 32.59 to 47.18, and the difference was not statistically significant with p=0.18. Across the validation datasets, performance remained consistent rather than drifting sharply from one hospital system to another.
That result is clinically consequential. A model does not need to outperform every existing bedside heuristic to have value if it can standardise risk assessment, reduce interpretive variation, and provide a reproducible estimate at a point when the fetal tracing is still early and potentially actionable. The CTG-derived variables likely matter because they encode the fetus’s compensatory response to labour stress: baseline rate, variability, accelerations, and decelerations each speak to different aspects of autonomic control and oxygen reserve. When those features are combined with pre-existing maternal and obstetric risk, the model can distinguish a fetus carrying little physiological margin from one likely to tolerate labour.
The sensitivity figures also set a realistic boundary. Fit4Labour is not a substitute for intrapartum clinical judgment, nor does it eliminate the need for escalation when the tracing deteriorates. It is a risk stratification tool, not a final arbiter.
Routine data may be the model’s greatest practical strength
Fit4Labour rests on information maternity teams already collect: routine obstetric history, the labour record, and the CTG trace itself. That makes implementation more plausible than a tool dependent on novel assays or extra bedside testing. Because the inputs are objective and reproducible, the output could be standardized across clinicians and sites, helping reduce the variability that has long complicated fetal monitoring.
The study’s retrospective design also means the model still needs prospective evaluation before it can be trusted in live labour wards as more than a validated prediction framework. Retrospective performance can show whether a signal exists and whether it generalises, but it cannot fully answer whether the model improves decisions, reduces avoidable harm, or changes intervention patterns without creating new problems such as unnecessary escalation. The next clinical hurdle is therefore not prediction alone, but whether early use of the score improves care without increasing low-value intervention.
For now, the evidence supports a clear proposition: early labour risk estimation for severe fetal compromise can be built from data that already exist at the bedside, and the resulting model performs consistently across three large maternity cohorts.