A novel machine-learning algorithm to screen for trisomy 21 in first-trimester singleton pregnancies.

Abstract

Background: Antenatal screening for Trisomy 21 (T21) in the UK is performed primarily in the first trimester. Nuchal Translucency (NT), gestational age, Free β-HCG and PAPP-A are used in combination, creating the 'combined' test. Multivariate Gaussian distribution models then determine the chance of T21 expressed as an odds ratio. This study investigates the use of machine-learning algorithms in the prediction of T21 in first-trimester singleton pregnancies and compares their performance to existing screening models.

Methods: A total of 86,354 anonymised, first trimester, singleton pregnancy screening cases, including 211 with T21, were used to train and test machine-learning models using adaptive boosting technology. Test case results were compared with pregnancy outcome data to assess performance.

Results: A machine-learning model was able to outperform current multivariate distribution models (McNemar's p = .006, AUC 0.978 vs 0.974). False positive rates were reduced from 3.82% to 2.28% (95% CI: 3.56-4.08 and 2.08-2.48 respectively) and overall screen positive rates were reduced from 4.00% to 2.48% (95% CI: 3.74-4.28 and 2.27-2.70 respectively).

Conclusions: Machine-learning algorithms offer demonstrable improvements to first-trimester singleton T21 screening without major changes to the UK programme. Larger datasets and improved outcome data would likely offer further increases in performance.