Artificial Intelligence to Improve Antenatal Detection of Congenital Heart Disease

Background: Congenital heart disease (CHD) is the most common group of fetal malformations, occurring in roughly 1% of pregnancies.

Cardiac views are part of the screening ultrasound offered to all pregnant women in the UK, but currently only around 54% of cases of severe CHD are diagnosed before birth.

This is a problem because antenatal detection has been shown to result in reduced mortality in children with CHD, and reduced rates of long-term morbidity such as neurodisability.

Machine learning (ML) is a form of artificial intelligence that has been shown to achieve human expert-level performance in some medical classification tasks.

Our group has developed ML tools that can automatically detect and store standard image planes from ultrasound data, and automatically measure biometry. Our early work has shown promising results in using ML to automatically detect abnormal fetal hearts.

I will build on this early work to explore AI-based methods to automatically identify fetuses that should be referred to specialist fetal cardiology services, and thus improve CHD detection rates.

Research question: How can deep learning algorithms be developed and utilised in the setting of fetal anomaly screening to improve antenatal detection rates of congenital heart disease?

Aims: To refine existing and develop new machine learning algorithms to automatically identify which fetuses need referral to specialist fetal cardiac services To show the clinical utility of the developed AI-based tools in a real-world setting, and demonstrate the potential of these tools to cause a paradigm shift in fetal anomaly screening worldwide Design and study methods: The proposed project is in two phases.

Phase 1: I will use existing retrospective imaging data to develop and test new algorithms to detect CHD.

I will explore different potential routes to this, including direct abnormality detection, use of cardiac biometry to identify markers of CHD, and combination of different models in an ensemble learning approach. The models will be combined into a clinically usable tool that will be tested in phase 2.

Phase 2: I will recruit 87 pregnant women, of which 29 will be carrying a fetus with CHD, and 58 volunteer sonographers, blinded to the CHD status of the fetuses. A full anomaly ultrasound scan will be performed between 18+0 and 23+6 weeks' gestation. Sonographers will be randomised to scanning either in the standard fashion or using the ML tool.

Each woman will be scanned using both techniques.

I will compare the detection rate of CHD between the two groups, the time taken for each scan, and the sonographer's subjective experience.

Anticipated impact of the project My goal is to develop a clinically useful tool that will increase the antenatal detection rate of CHD.

If the proposed system can demonstrate this, then it is likely to translate into an improved postnatal condition of affected infants.

In turn this is likely to reduce postnatal mortality, and the incidence of neurological injury, which may reduce some of the long-term morbidity associated with CHD.