AI may improve accuracy of 'combined' prenatal test for Down's syndrome

A new approach to high-chance prenatal Down's syndrome screening, involving AI, could improve the predictive accuracy of the 'combined' test for Down's syndrome.

A proof-of-concept clinical trial at Royal Bolton Hospital used machine learning, to decipher patterns within combined test data that traditional statistical methods might miss. Compared with standard approaches, this AI-driven approach more accurately predicted high-chance results, potentially reducing the number of women referred for invasive procedures such as amniocentesis and chorionic villus sampling (CVS).

'With the rapid growth of AI, it felt like an opportunity to use the technology to interpret and identify patterns in numbers that we might not be able to see,' said Jamie Osborne – principal clinical scientist at Bolton NHS Foundation Trust and lead author of the research, which was published in the Journal of Obstetrics and Gynaecology. 'The more data we can generate, the more accurate the testing and theory will become.'

Bolton NHS Trust encompasses the second largest antenatal screening laboratory in the UK. Staff there hypothesised that modern machine learning techniques could prove to be a cheaper and more accurate first-line screening tool in first-trimester singleton pregnancies.

At around 12 weeks' gestation, pregnant women in England are routinely offered a combined test which integrates ultrasound results (fetal nuchal translucency), gestational age and blood biomarker levels to estimate the chance of trisomy 21 (the cause of Down's syndrome). Currently, these tests provide an odds ratio which determines whether further testing – such as NIPT, amniocentesis or CVS – will be offered.

Rather than employing the better known 'deep learning' method of machine learning, the staff at Bolton NHS Trust turned to the 'adaptive boosting' method, known as 'Adaboost' for short. This is a 'random forest' method in which decision trees are built sequentially, and each new tree focuses on cases that were previously misclassified. When a new case is evaluated, all trees 'vote' on whether the chance of Down's syndrome is high or low. These votes are then combined, with greater weight given to votes from trees that performed better during training.

The team sampled 86,354 anonymised first-trimester cases of singleton pregnancy from the Newcastle upon Tyne Hospitals NHS Trust. This was supplemented with synthetic data in order to address severe imbalances in screening datasets, which would otherwise have made a machine learning approach impractical.

The study showed that the AI-driven approach outperformed current methods, reducing false positive rates from 3.8 percent to 2.3 percent. Further evidence for the merits of such an approach comes from a recent Turkish study, which suggests that more accurate combined tests can improve early diagnosis and thereby reduce unnecessary invasive procedures.

'This has given us the confidence that we can use biochemical data to improve accuracy, health outcomes and services,' said Osborne. 'This may open doors in the future to use biochemical tests for the screening of diseases such as pancreatic cancer, which is difficult to diagnose and has low survival rates.'