[Paper Review] Using Interpretable Machine Learning to Predict Maternal and Fetal Outcomes

Tomas M. Bosschieter, Zifei Xu+6 more|arXiv (Cornell University)|Jul 12, 2022

Statistical Methods in Epidemiology4 citations

TL;DR

This study uses Explainable Boosting Machines (EBMs), a glass-box interpretable machine learning model, to predict maternal and fetal complications—severe maternal morbidity (SMM), shoulder dystocia, and preterm preeclampsia—using a large, real-world obstetric dataset of 222,001 de-identified patients. EBMs achieve predictive performance on par with black-box models like XGBoost and deep neural networks while revealing counterintuitive risk factors, such as time from full cervical dilation to delivery and maternal height, challenging traditional clinical assumptions.

ABSTRACT

Most pregnancies and births result in a good outcome, but complications are not uncommon and when they do occur, they can be associated with serious implications for mothers and babies. Predictive modeling has the potential to improve outcomes through better understanding of risk factors, heightened surveillance, and more timely and appropriate interventions, thereby helping obstetricians deliver better care. For three types of complications we identify and study the most important risk factors using Explainable Boosting Machine (EBM), a glass box model, in order to gain intelligibility: (i) Severe Maternal Morbidity (SMM), (ii) shoulder dystocia, and (iii) preterm preeclampsia. While using the interpretability of EBM's to reveal surprising insights into the features contributing to risk, our experiments show EBMs match the accuracy of other black-box ML methods such as deep neural nets and random forests.

Motivation & Objective

• To identify and quantify the most important risk factors for three serious obstetric complications: SMM, shoulder dystocia, and preterm preeclampsia.
• To evaluate the performance of interpretable machine learning models, specifically Explainable Boosting Machines (EBMs), in predicting these outcomes with high accuracy and transparency.
• To challenge and revise traditional clinical assumptions about risk factors by uncovering surprising, data-driven relationships through model interpretability.
• To demonstrate that EBMs can outperform standard logistic regression in predictive performance while maintaining full interpretability, making them suitable for clinical decision support.

Proposed method

• The study employs Explainable Boosting Machines (EBMs), a type of Generalized Additive Model (GAM) that uses boosting to learn univariate shape functions for each feature, enabling global and local interpretability.
• EBMs are trained on a de-identified clinical dataset from 20 U.S. hospitals, containing 408 features and 222,001 patient records, with outcomes defined by clinical diagnosis codes.
• Features are selected based on clinical plausibility and temporal precedence—only variables measurable before the outcome can be used in prediction.
• The model automatically detects and incorporates important pairwise interaction terms to improve predictive accuracy while preserving interpretability.
• Model performance is evaluated using AUROC and calibration curves, comparing EBMs to XGBoost, random forests, deep neural networks, and logistic regression.
• Interpretability is visualized through shape functions that show the marginal effect of each feature on the predicted risk, with error bars indicating uncertainty from outer-bagging.

Experimental results

Research questions

• RQ1 Which clinical features are the most important predictors of severe maternal morbidity (SMM), shoulder dystocia, and preterm preeclampsia according to an interpretable machine learning model?
• RQ2 How does the performance of Explainable Boosting Machines (EBMs) compare to black-box models like XGBoost and deep neural networks in predicting obstetric complications?
• RQ3 What surprising or non-intuitive risk factor relationships does the EBM model reveal that challenge traditional clinical assumptions?
• RQ4 To what extent do demographic factors like race and ethnicity contribute to risk, and how are these effects mediated by clinical and biological variables?

Key findings

• EBMs achieved AUROC scores on par with XGBoost, random forests, deep neural networks, and logistic regression, demonstrating strong predictive performance while remaining fully interpretable.
• The most important risk factor for SMM was preeclampsia or gestational hypertension, followed by time from full cervical dilation to delivery and history of cesarean section.
• For shoulder dystocia, baby’s weight was the top contributor, but time from membrane rupture to delivery and maternal height were also highly significant, challenging the primacy of fetal size and maternal diabetes.
• For preterm preeclampsia, maternal BMI, number of previous stillbirths, and pre-pregnancy hypertension were the top three risk factors, with maternal age over 44 showing a strong contribution.
• The model revealed a clinically surprising near-linear relationship between fetal weight (3250g–4250g) and risk of shoulder dystocia, suggesting a threshold effect not previously emphasized.
• Non-White race and Hispanic/Latino ethnicity were associated with higher risks for both SMM and shoulder dystocia, though these associations may be mediated by access to care and other correlated factors.