Early prediction of pregnancy disorders with machine learning guided Raman spectroscopy, and metabolomics.
Date
2022-08
Authors
Synan, Lilly C
Major Professor
Advisor
Bardhan, Rizia
Mallapragada, Surya
Kohut, Marian
Committee Member
Journal Title
Journal ISSN
Volume Title
Publisher
Altmetrics
Research Projects
Journal Issue
Series
Department
Chemical and Biological Engineering
Abstract
Pregnancy abnormalities are often discovered during the second or third trimester with current screening tools available in clinics. Treatments are rendered less effective when there is less time to prepare for delivery, and earlier prediction measures are needed to ensure safer pregnancy outcomes. Preterm labor (PTL) has a perinatal mortality rate of 10.7%, and gestational diabetes mellitus (GDM) is the most common global pregnancy abnormality. Raman spectroscopy (RS) can be used as a metabolic profiling technique for clinical samples in the first trimester to detect at-risk pregnancies. T-stochastic neighbor embedding (tSNE) is a machine learning technique that was used show visual separation of both disorders, while partial least squares-discriminant analysis (PLS-DA) was also applied to GDM data to explore potential for predictive modeling. AUC-Roc analysis successfully classified the PTL group with an AUC = 0.831 and a 95% confidence interval of 0.515-1, while GDM had an AUC = 0.99 and a 95% confidence interval of 1-1. Mass spectrometry was performed on a subset of patient samples to confirm Raman findings for each group. PTL was differentiated from healthy with 13 statistically significant metabolites, while GDM had 17. Next, the biomarkers were combined with clinical data, depicting even greater classification potential within tSNE and PLS-DA. These findings provide the foundation for modeling raman and clinical data so that it can predict the likelihood of PTL or GDM developing in new pregnancies in future studies.