Computers have been trained to analyse the CTG and interpret whether it is normal or not. A recent literature review has explored the current state of play for computer interpretation of the CTG (M’Barek, et al., 2022). There have been previous reviews of this literature but it is a rapidly evolving space. M’Barek and colleagues bring us back up to date. This post summarises the key findings of their literature review.
Guideline based computer analysis
A strength of this review is the recognition that not all systems providing computer analysis of the CTG are the same. The first approaches to developing computer systems to analyse CTG recordings taught computers to recognise features of the trace on the basis of existing guidelines. The FIGO guideline forms the interpretive code for the Dawes-Redman system developed in the 1980s, and also the Portuguese Omniview-SisPorto system. A more recent addition to this type of system is OxSys – here the focus is on deceleration area rather than the full set of CTG features in the FIGO guideline.
Machine learning plus guideline based analysis
The problem with guideline based systems is that if guideline defined CTG abnormalities perform poorly and computers are trained to detect only the guideline defined abnormalities, then perinatal outcomes are unlikely to improve to any great extent. The next advance in computer analysis has been to offer a large database of CTG tracings, where perinatal outcomes are known, and let the computer get to work. These systems still make use of a guideline based approach, but add things like spectral and multi-fractal analysis to identify other potential markers of developing hypoxia.
Deep learning analysis
Others have let go of guideline based analysis entirely, and use an approach called deep learning. Raw CTG data from cohorts with a known outcome is presented to the computer for analysis. Most of these training systems have used cord blood acidosis as the outcome they are attempting to detect, while others have used low Apgar scores. Database sizes have ranged from only 36 to as many as 38,000 CTGs.
Does it work?
To date the only randomised controlled trials investigating the impact of computer analysis rather than only using clinician interpretation of the CTG have made use of guideline based systems. These three trials have been included in two meta-analyses, with both showing computer analysis offered no perinatal benefits compared to clinician interpretation. Machine learning and deep learning approaches have not yet been assessed for their potential impact on clinical outcomes.
Where to next?
The original goal for intrapartum CTG monitoring was to prevent death or permanent neurological injury of the fetus, while avoiding unnecessary intervention for the labouring woman. It seems likely that researchers will continue to attempt to achieve that outcome with the tools available, and machine learning is the newest tool to be added to the mix. The authors of this review raise the following points for consideration:
- Will clinicians trust the recommendation given by a computer to intervene, or not, when the visual interpretation of the CTG by the clinician is at odds with what the computer can see? Spectral and multi-fractal analysis looks for changes that are not recognisable to human eyes. It will require a leap of faith to rely on a computer generated recommendation.
- How do we evaluate new systems? So far, training of computers has been done on retrospective collections of CTGs. How do we safely test these systems in clinical practice, and what do we compare them with? Clinician interpretation of the CTG? Intermittent auscultation? The low incidence of the outcomes of interest has always plagued fetal heart rate monitoring research, requiring randomised controlled trials to recruit very large populations.
- What are the outcomes we should be training systems to prevent? Cord blood acidosis and Apgar scores are proxy measures for clinical measures, but they don’t consistently correlate with outcomes such as neonatal mortality or neurodevelopment delay. Longer term outcomes are important and have been understudied in research to date. Will this continue, or will attempts be made to aim for changes in outcomes like cerebral palsy?
My final thoughts
It it important to keep in mind that the databases used to train computers in CTG interpretation come from clinical populations. Each of these women was receiving clinical care, in a complex maternity care system, with a strong focus on achieving good outcomes for the fetus and newborn. Other common clinical practices, such as early cord clamping, coached pushing, oxytocin use, and epidural analgesia, have the potential to confound the research, and I am not confident that these known and unknown confounders have been, or even can be, managed.
At the end of this very expensive and time consuming process, if researchers are show intrapartum monitoring of the fetal heart rate DOES NOT prevent fetal harm from hypoxia, will we finally let go of it? Or will we replace computer analysis of the heart rate with whatever the next big technology advance is at the time? Will a deep learning approach offers potential useful insights into fetal physiology that traditional approaches to studying physiology can’t provide us?
Is it possible to develop and verify approaches to computer interpretation of the fetal heart rate in ways that don’t reproduce and reinforce patriarchal values? I suspect that Audre Lorde’s words about not being able to use the Master’s tool to destroy the Master’s house apply here. Computer interpretation originated from a technocentric world view. It seems destined to be applied in ways that reinforce the core beliefs that women’s bodies are risky and can’t be trusted. Is that the future we are aiming for in maternity care?
References
Ben M’Barek, I., Jauvion, G., & Ceccaldi, P. F. (2022, Dec 20). Computerized cardiotocography analysis during labor – A state-of-the-art review. Acta Obstetrica et Gynecologica Scandinavica, in press. https://doi.org/10.1111/aogs.14498
Categories: CTG, EFM, Perinatal brain injury, Perinatal mortality
Tags: computerised interpretation
Birth Small Talk 