ABSTRACT
BACKGROUND: Historically, noninvasive techniques are only able to identify chromosomal anomalies that accounted for <50% of all congenital defects; the other congenital defects are diagnosed via ultrasound evaluations in the later stages of pregnancy. Metabolomic analysis may provide an important improvement, potentially addressing the need for novel noninvasive and multicomprehensive early prenatal screening tools. A growing body of evidence outlines notable metabolic alterations in different biofluids derived from pregnant women carrying fetuses with malformations, suggesting that such an approach may allow the discovery of biomarkers common to most fetal malformations. In addition, metabolomic investigations are inexpensive, fast, and risk-free and often generate high performance screening tests that may allow early detection of a given pathology. OBJECTIVE: This study aimed to evaluate the diagnostic accuracy of an ensemble machine learning model based on maternal serum metabolomic signatures for detecting fetal malformations, including both chromosomal anomalies and structural defects. STUDY DESIGN: This was a multicenter observational retrospective study that included 2 different arms. In the first arm, a total of 654 Italian pregnant women (334 cases with fetuses with malformations and 320 controls with normal developing fetuses) were enrolled and used to train an ensemble machine learning classification model based on serum metabolomics profiles. In the second arm, serum samples obtained from 1935 participants of the New Zealand Screening for Pregnancy Endpoints study were blindly analyzed and used as a validation cohort. Untargeted metabolomics analysis was performed via gas chromatography-mass spectrometry. Of note, 9 individual machine learning classification models were built and optimized via cross-validation (partial least squares-discriminant analysis, linear discriminant analysis, naïve Bayes, decision tree, random forest, k-nearest neighbor, artificial neural network, support vector machine, and logistic regression). An ensemble of the models was developed according to a voting scheme statistically weighted by the cross-validation accuracy and classification confidence of the individual models. This ensemble machine learning system was used to screen the validation cohort. RESULTS: Significant metabolic differences were detected in women carrying fetuses with malformations, who exhibited lower amounts of palmitic, myristic, and stearic acids; N-α-acetyllysine; glucose; L-acetylcarnitine; fructose; para-cresol; and xylose and higher levels of serine, alanine, urea, progesterone, and valine (P<.05), compared with controls. When applied to the validation cohort, the screening test showed a 99.4%±0.6% accuracy (specificity of 99.9%±0.1% [1892 of 1894 controls correctly identified] with a sensitivity of 78%±6% [32 of 41 fetal malformations correctly identified]). CONCLUSION: This study provided clinical validation of a metabolomics-based prenatal screening test to detect the presence of congenital defects. Further investigations are needed to enable the identification of the type of malformation and to confirm these findings on even larger study populations.
