Probabilistic machine learning methods for Down syndrome prenatal screening: unified first and second trimester markers multivariate bayesian risk model versus current contingent sequential strategy

The aim of the study is to compare performance of a multivariate bayesian model to estimate Down syndrome (DS) risk from a dataset containing biochemical and ultrasound parameters used in first and second-trimester prenatal screening with current contingent sequential screening strategy in the diagnostic of cases missed in first-trimester screening. Methods: observational study in a retrospective way of collecting data from pregnant women who attended the Lluís Alcanyís, Ribera de Alzira y Verge dels Lliris de Alcoy hospitals in Valencia -Spain- for a period of 10 years (2010-2020). Multivariate normal distributions for affected and unaffected fetuses, specified by the means and the covariance matrix of biochemical and ultrasound parameters were set up. Area under the curve (AUC) of ROC (receiver operating characteristic) for classification of DS missed in first trimester screening was the figure to compare by DeLong test the diagnostic performance of quadratic discriminant analysis from this multivariate Bayesian model to current Naïve-Bayes strategy. Results: no significant differences were found between both strategies. Other than second-trimester biometric variables between them, no relevant correlation between the rest of markers was found. Conclusion: accuracy of other machine learning strategies should be investigated, namely artificial neural network, to improve the identification of first trimester undetected cases.