The biggest weakness of fetal heart rate monitoring is not the monitoring itself but its interpretation: the same trace is read differently by different clinicians. Artificial intelligence promises to reduce this subjectivity — but the largest randomized trial showed that computerized decision support did not improve clinical outcome. We examine the evidence as it stands.
⚕️ Target audience: This content is aimed at healthcare professionals and is for educational/current-information purposes. AI tools offer clinical decision support; It does not replace clinician responsibility and approved clinical workflows.
Cardiotocography (CTG — simultaneous recording of fetal heart rate and uterine contractions; called non-stress test/NST when used antepartum) has been at the center of obstetric care for half a century. Continuous electronic fetal monitoring (EFM) has become the standard method for assessing fetal well-being in labor. But the paradox of the method is this: it is easy to record the trace, but difficult to interpret it correctly. CTG signals are complex and inconsistent as evaluation based on visual interpretation remains subjective. This is where artificial intelligence (AI) enters, with the promise of reducing subjectivity and increasing interpretation consistency.
The best-documented problem limiting the clinical value of CTG is low interobserver and intraobserver agreement—that is, different clinicians may classify the same trace differently, or even the same clinician at different times. In a study testing the reliability of the NICHD three-tier (3-Tier) fetal heart rate classification system, interobserver agreement was found to be only moderate (kappa 0.45). More striking was that in the most critical category—pathological (Category III) tracings—the agreement was poor (kappa 0.0); this discrepancy was largely due to disagreement on the distinction between “none” and “minimal” variability (Blackwell et al., 2011).
A similar picture is also seen in European data. In a study using FIGO 2015 classification criteria, moderate agreement was achieved in classifying intrapartum tracings as "normal", while no consensus was reached in the "suspicious" and "pathological" categories (Amadori et al., 2022). The main justification for seeking automation is that interpretation becomes most unreliable at the point where it is most clinically important—abnormal traces. This variability is not an academic detail: it can lead to unnecessary interventions (cesarean section, operative delivery) and may overlook the real fetal danger.
Long before the current wave of artificial intelligence, rule-based computational analysis attempted to solve this problem. Developed in the early 1990s, the Dawes-Redman system (commercially in widespread use as Oxford Sonicaid/Oxford System) is a standardized automated analysis method that evaluates antepartum CTG signals with empirically derived criteria. The system quantifies a parameter that cannot be reliably measured by visual interpretation—short-term variability (STV—the subtle millisecond-level variability between fetal heartbeats)—and provides an objective measure for assessing fetal well-being.
Updated iteratively over three decades and containing more than 100,000 pregnancy data, the Dawes-Redman system enables robust analysis of signal properties and is an established tool in antepartum monitoring—especially fetal growth restriction monitoring (Jones, Cooke, Vatish, & Redman, 2022). An important conceptual note: Dawes-Redman is a "rule-based" expert system; It is methodologically different from modern machine learning/deep learning approaches that learn patterns from data. Nevertheless, it remains the most widely accepted computerized CTG approach in clinical practice today.
In recent years, the research focus has shifted to data-driven deep learning (deep learning - a sub-branch of machine learning that extracts patterns from raw data with multilayer artificial neural networks) methods that learn directly from the signal. These models aim to automatically detect fetal compromise (risk of fetal asphyxia/acidosis) by processing fetal heart rate time series. A comprehensive scoping review dated 2023 mapped 40 different studies examining at least one algorithm classifying CTG traces; emphasized that these approaches have the potential to reduce observer variability, but none have gained wide acceptance in clinical practice and more curated and labeled data are needed to achieve a validated decision support system (Aeberhard et al., 2023).
The most critical weakness of this field is the lack of generalizability and external validation. Models are often developed on single and unbalanced data sets; Their performance may decrease when transferred to different populations and devices. A 2025 study evaluated six deep learning models cross-database on a proprietary dataset (9,887 records) and the open-access CTU-UHB dataset (552 records). ResNet architecture showed the strongest performance in both databases; Furthermore, the model's class-activation maps of high-contributing signaling regions overlapped with clinical knowledge of the compromised fetal patterns—that is, what the model was "looking at" was interpretable. While the study proposes a standardized workflow for comparison of future methods, its main message is that new methods should not be introduced into clinical practice without demonstrating generalizability and interpretability (Mendis et al., 2025).
📊 How should performance figures be read? In current multicenter studies, deep learning models report AUC (area under the curve; measure of discriminatory power) around 0.86 in external test sets (e.g. Park et al., 2025). These figures are promising, but are retrospective and based largely on single-country data; There is no prospective evidence to show whether it improves clinical outcome (neonatal health). High AUC does not imply clinical benefit.
The strongest justification for avoiding hyperbole in this area is the largest randomized controlled trial that has directly tested the issue. The INFANT study (INFANT Collaborative Group, Lancet 2017) tested whether adding decision support software to CTG interpretation affected the number of poor neonatal outcomes in women who underwent continuous EFM in labor. The study is massive: 47,062 women were randomized across 24 maternity units in the UK and Ireland, with 46,042 analysed.
The conclusion is clear and negative: no difference was detected between groups in the primary composite adverse neonatal outcome (intrapartum stillbirth/early neonatal death, neonatal encephalopathy, or neonatal unit admission with evidence of compromise at birth)—0.7% (172 infants) in the decision-support group vs. 0.7% (171 infants) in the no-decision-support group; adjusted risk ratio (RR) 1.01, 95% confidence interval (CI) 0.82–1.25. There was also no significant difference in neurodevelopmental assessment at age two. The study's own conclusion is explicit: "The use of computerized CTG interpretation in women undergoing continuous electronic fetal monitoring in labor does not improve clinical outcomes for mothers or infants."
The detailed health technology assessment report (Brocklehurst et al., Health Technol Assess 2018) adds that decision support changed practice but did not change outcome: fetal blood sampling increased and the rate of repetitive stimulation decreased in the decision support arm; However, there was no difference in healthcare resource use and total cost. That is, the system influenced clinician behavior, but this change in behavior did not translate into better neonatal outcome. The lasting lesson INFANT leaves is that even if the technology produces a more accurate alert, the real bottleneck is often the timing and appropriateness of the clinical response to CTG abnormalities—which is still highlighted as an urgent need for education.
⚠️ Artificial intelligence does not always surpass humans. In an experimental comparison from 2025 (Miyata et al.), human expert judgment achieved the highest AUC in predicting fetal asphyxia (0.693), surpassing machine learning (0.514) and deep learning (0.524). Although deep learning can catch some cases that humans miss and increase sensitivity, it cannot surpass the accuracy of human judgment alone. Interestingly, when humans and AI were combined, specificity increased (from 91.9% to 98.0%)—so the real contribution of AI may not be to “replace humans” but to complement it by reducing false-positives. This is a single small-scale study and should not be generalized, but it is a solid reminder against exaggeration.
The practical picture today is as follows: In antepartum follow-up, Dawes-Redman type computerized analysis is an established and useful objective tool, especially with STV measurement, and is used in high-risk follow-up. On the other hand, deep learning-based decision support systems in labor are still in the research phase; There is no intrapartum AI system validated for routine clinical use that has been prospectively proven to improve outcomes. The logical stance is to position AI not as a decision maker that replaces the clinician, but as an assistant that directs attention and increases consistency.
The questions that remain open are: In which patient population will real benefit be achieved? Do the models generalize across different centers? Most importantly — does a more accurate alert translate into a better clinical response and better neonatal outcome? What will answer these questions are well-designed prospective/randomized studies and compliance with transparent reporting standards (e.g. TRIPOD-AI, DECIDE-AI), rather than AUC tables. Until that day, the safest approach is to evaluate AI output along with critical clinical judgment and invest in clinical training to respond appropriately to CTG abnormalities in a timely manner.
Last updated: . Sources have been confirmed with their publication dates and current versions.