Feasibility of Predicting Rare Adverse Outcomes of Pregnancy Using Machine Learning: Can an Algorithm Use Fetal to Placental Weight Ratios to Predict Which Pregnancies End in Stillbirth? 公开

Abstract

Stillbirth, a rare adverse pregnancy outcome, continues to negatively impact families worldwide. However, the etiology of many stillbirths is uncertain, and novel prevention strategies are needed. Some evidence suggests abnormal fetal to placental weight ratios are  associated with increased risk of stillbirth. Although a method exists to estimate placental volume (EPV) during pregnancy, this tool has not yet been studied for stillbirth prevention. Given the challenges of conducting a prospective study for this rare outcome, we used machine learning to develop algorithms to predict stillbirth and evaluate the role of fetal to placental weight ratios in predictive accuracy.

We used Medical Birth Registry data from Norway for approximately 1 million women with singleton pregnancies ( > 24 weeks). We created a dataset with stillbirth as the outcome, fetal to placental weight ratio as the main covariate of interest and multiple additional covariates to potentially improve predictive accuracy. We used a stacking algorithm, Super Learner, that combines several standard regression and machine learning algorithms into one, via a 5-fold cross validation. We used a 70:30 train:test data split to fit 3 models for predicting stillbirth: one with basic covariates only, one with EPV only, and a third with all variables included. Specificities and sensitivities were calculated, using the test dataset, to develop receiver operating characteristic (ROC) curves with which we assessed the area under the curve (AUC). We evaluated concordance measures for each algorithm.

While results are pending application of our programs to the Norway Medical Birth Registry data by our Norway colleagues, we expect to receive outputs from the three models that estimate predictive accuracy and intend to evaluate these results in two dimensions: absolute performance (how well each algorithm predicts stillbirth) and relative performance (how each algorithm compares to the other).

Overall, stillbirth remains a pregnancy issue requiring greater awareness and research efforts. Hence, prediction studies, such as this, can provide insight into how future studies may be constructed to predict stillbirth, and other rare adverse health outcomes.

Table of Contents

Introduction 1

Methods 2

Statistical Analysis 3

Results 5

Discussion 10

References 15

Appendix A 18

About this Master's Thesis

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Rollins School of Public Health
Department	Epidemiology
Degree	M.P.H.
Submission	Master's Thesis
Language	English
Research Field	Health Sciences, Public Health Health Sciences, Epidemiology
关键词	Super Learner receiver operating characteristic curve prediction fetal to placental weight ratio area under the curve machine learning stillbirth estimated placental volume
Committee Chair / Thesis Advisor	Naimi, Ashley, Emory University Christiansen-Lindquist, Lauren, Emory University

最新修改

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Feasibility of Predicting Rare Adverse Outcomes of Pregnancy Using Machine Learning: Can an Algorithm Use Fetal to Placental Weight Ratios to Predict Which Pregnancies End in Stillbirth? ()	2023-04-26 17:01:41 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions