The use of machine learning in obstetric ultrasound: a mixed methods study examining clinical value and acceptability
Bradburn EH.
Background: Despite its importance, accurate gestational age (GA) estimation remains a significant challenge, especially in low-and middle-income countries. The gold standard, which is to measure the fetal crown rump length between 11-14 weeks, is often not achievable in these settings as women present late in pregnancy or do not have access to trained ultrasonologists. Later GA assessment is currently estimated by fetal biometry, but as gestation advances this becomes less accurate as variation in fetal size increases. In this thesis I explore the hypothesis that alternative methods of GA estimation, that require less ultrasound training, are available and acceptable. I assess liquid biomarkers; and machine learning (ML) methods of GA estimation based on ultrasound images without fetal biometry. With the latter method showing promise, I explore the opinions of healthcare workers and pregnant women to the introduction of ML within maternity care. Objectives: Mixed methods were used to (1) examine which maternal or newborn biomarkers can be used to estimate GA (2) co-design a ML method that uses ultrasound image analysis to estimate GA (3) validate and establish the performance of this ML method in a large international cohort of pregnant women (4) understand trust in ML technologies and potential barriers to their incorporation in maternity care amongst healthcare professionals and pregnant women. Methods: (1) A systematic review and meta-analysis was conducted to establish which biomarkers are able to estimate GA. (2) & (3) Second and third trimester ultrasound images from the international INTERGROWTH-21st were used to develop and internally validate an algorithm to estimate GA based on image analysis. This was then externally validated using ultrasound images from INTERBIO-21st. All women had previous ground-truth GA estimation as a reference standard. (4) A prospective, multi-centre, qualitative study in Nigeria, Rwanda, UK and Canada was conducted using semi-structured interviews with healthcare workers and pregnant women. Results: (1) No prenatal maternal biomarker that accurately predicts GA was found. In newborns, metabolomic profiling from blood spot produced the most accurate estimate of GA (at birth) with a pooled RMSE of 1.05 weeks, with a slight drop in accuracy in small for GA (SGA) and preterm infants. For (2) & (3), a MultiPlane algorithm outperformed biometry-based methods of GA estimation in external validation, with a mean absolute error (MAE) of 3 and 4 days in the second and third trimesters, respectively. Sub-group analysis showed that this was also more accurate than current methods of GA estimation in babies born SGA and LGA, and in women with raised BMI. (4) Healthcare workers and pregnant women welcome the use of AI in maternity care and recognise its potential to improve healthcare. However, this work identified important barriers that must be addressed for successful implementation. Conclusion: While postnatal metabolomic profiling is able to estimate GA, this is not available during pregnancy, not a scalable solution and inferior to ultrasound dating. A ML method of image analysis improved on current methods of GA estimation using biometry, and performs well in SGA and LGA babies. As it does not require training for detailed biometry, it has important potential in LMIC settings. Healthcare workers and pregnant women feel positive towards ML within maternity care, however important barriers must be addressed.