Next Generation Assessment of Fetal Wellbeing using Artificial Intelligence

Lead Research Organisation: University of Oxford
Department Name: Women s and Reproductive Health


Electronic Fetal Heart monitoring (also called CTG) is the measurement of the fetal heart rate using probes that are placed on the mother's abdomen. It is the commonest test of fetal wellbeing worldwide (>200M tests per year are performed). It is used to try and assess how healthy the baby is and produces a readout that is very complex; surprisingly this readout is usually analyzed visually (by eye) and the clinician performing this analysis will use this to justify whether a baby needs to be delivered or not. There is a substantial amount of published data that shows this visual assessment is extremely poor and groups of clinicians disagree about a CTG and even the same doctor can interpret a CTG differently on different days. This means that some babies are delivered too soon and many sick babies are delivered too late - both create major problems for the babies, their parents, the NHS and society. CTG can be performed in pregnancy before labour or during labour. Most stillbirths occur before labour (>80%) and we have focused on this area.
Many groups (including ourselves) have tried to standardize assessment of the CTG readouts using rudimentary computerised assessment. These systems certainly reduce the disagreements between clinicians but can't account for the multiple factors that make a CTG normal or abnormal (e.g. no systems account for the gestational age of a baby - e.g. treating a 28 week baby the same as a 38 week baby. These systems don't incorporate maternal or fetal disease into the analysis and none of these systems can tell clinicians what will happen to the baby in the coming days or weeks.

We have brought together a team with expertise in CTG and artificial intelligence to deliver next generation assessment of the fetus.

We will develop a suite of artificial intelligence based machine-learning models to revolutionise antepartum CTG analysis. Recent advances in deep-neural-networks (DNN) enable advanced analysis and identification of novel features within these complex signal patterns which we can exploit in conjunction with detailed maternal and fetal clinical outcomes to generate high fidelity diagnostic and prognostic tools. At our disposal is a unique unrivaled database of >165,000 fully classified CTG signals with associated maternal and neonatal outcome data from >56,000 pregnancies. Leveraging these data, we will develop artificial intelligence based technologies specific to the unique context of the mother and the fetus (at any gestational age). We have proven experience of analysing large datasets using these AI based tools. We have built into our plans the ability to validate our findings with prospective data from Oxford and Melbourne. The potential health benefits are substantial.

Our work streams will allow us to generate clean data, generate tools that will allow our AI solution to be used on any CTG from any manufacturer. We will incorporate gestational age, maternal and fetal disease status and provide clinicians with a precise risk assessment of the fetus that will significantly improve the way we care for babies in the UK and beyond.

Technical Summary

Cardiotocography (CTG) is the non-invasive measurement of the fetal heart rate (FHR) and uterine activity. CTG is the gold standard test of fetal wellbeing worldwide (>1M/year UK; >200M/year globally). 85-98% of pregnancies will undergo at least one CTG, frequently more. However, CTG patterns are complex and difficult to interpret, leading to unnecessary interventions and missed opportunities. This is important in the context of stillbirth (8 families/day UK). Antenatal CTG (aCTG: 26 weeks until labour) is where the vast majority of severe adverse fetal outcomes occur (e.g. 90% of stillbirths). The complex signals in aCTG are overwhelmingly assessed visually by clinicians. As such, it suffers from subjectivity, low accuracy and poor reproducibility. In the worst cases, clinicians either fail to deliver at-risk babies (resulting in stillbirth) or intervene too early, exposing the baby to avoidable risk. Complications from pre-term birth are the leading cause of mortality in children <5 yrs. Improving CTG analysis will substantially help solve this problem as recommended by NHS England. We have built an unrivalled database (10x larger than any others) of aCTGs (>165,000) and comprehensive maternal and fetal outcomes (>56,000 pregnancies). This is perfect for ML. In collaboration with Computer Science, Oxford University and the Alan Turing Institute, we have developed preliminary algorithms that significantly outperform current approaches (AUC 0.82-0.91 vs. 0.63). In this project, we will A) further develop and refine our preliminary ML-driven data processing and risk identification models, and B) Validate these models using external validation data. This work will pave the way for a clinical trial and displace current, insufficient approaches to CTG analysis.


10 25 50