A Comprehensive and Bias-Free Machine Learning Approach for Risk Prediction of Preeclampsia with Severe Features in a Nulliparous Study Cohort

Abstract

Objective: Preeclampsia is one of the leading causes of maternal morbidity, with consequences during and after pregnancy. Because of its diverse clinical presentation, preeclampsia is an adverse pregnancy outcome that is uniquely challenging to predict and manage. In this paper, we developed machine learning models that predict the onset of preeclampsia with severe features or eclampsia at discrete time points in a nulliparous pregnant study cohort.

Materials and methods: The prospective study cohort to which we applied machine learning is the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-be (nuMoM2b) study, which contains information from eight clinical sites across the US. Maternal serum samples were collected for 1,857 individuals between the first and second trimesters. These patients with serum samples collected are selected as the final cohort.

Results: Our prediction models achieved an AUROC of 0.72 (95% CI, 0.69-0.76), 0.75 (95% CI, 0.71-0.79), and 0.77 (95% CI, 0.74-0.80), respectively, for the three visits. Our initial models were biased toward non-Hispanic black participants with a high predictive equality ratio of 1.31. We corrected this bias and reduced this ratio to 1.14. The top features stress the importance of using several tests, particularly for biomarkers and ultrasound measurements. Placental analytes were strong predictors for screening for the early onset of preeclampsia with severe features in the first two trimesters.

Conclusion: Experiments suggest that it is possible to create racial bias-free early screening models to predict the patients at risk of developing preeclampsia with severe features or eclampsia nulliparous pregnant study cohort.

Figure 1. Data process timeline

This figure shows the gestational weeks at each visit. For each visit, the number of features at that visit is listed and the category of new feature included is also shown. ¹ Specific medical condition can be found in Table Supplement 5 ² Specific placental analytes can be found in Table Supplement 3

Figure 2. Final study cohort selection process.

Out of the participants from the placental analytes sub-study, we excluded participants with conditions such as chronic hypertension, mild preeclampsia, and missing label for preeclampsia to focus on the participants that are most at risk.

Figure 3. The training process of PEPrML pipeline.

Samples were balanced for train and test sets. 5-fold grid search cross-validation was used to select the hyperparameters for each trial. We repeated 100 trials and recorded the results.

Figure 4. sPE+E vs NPH and early sPE vs late sPE+E model performance.

a) Average AUC for 100 trials per visit for 4 classifiers. b) RF classifier has best performance across visits for both comparisons. The ROC curve demonstrated the tradeoff between the true positive rate versus false positive rate. This summarizes the results for 100 trials.

Figure 6. Fairness check for sPE+E vs NPH mode

The threshold set based on the four-fifth rule are 0.8 and 1.25. Ceribus Paribus plot was used to adjust prediction threshold for the Black population.