Delfina was honored to participate in the Decoding Maternal Morbidity Data Challenge run by the NICHD, and to have our submission selected for awards for both innovation as well as considering health disparities for disadvantaged groups. We set up a collaboration between members of our Medical Advisory Board, our data science team, and our software engineering team to explore this dataset and come up with a meaningful approach. For our submission, we trained a model to predict hypertensive disorders of pregnancy at term on the nuMoM2b dataset, which is currently deployed here.
Hypertensive disorders of pregnancy (HDP), including gestational hypertension, preeclampsia, and eclampsia, are a significant cause of maternal morbidity and mortality in the United States. The American College of Obstetricians and Gynecologists has estimated that HDP complicates up to 10% of pregnancies. Risk stratification for HDP development currently relies on maternal history and risk factors analysis to identify high risk patients for closer monitoring and intervention. This overall has limited predictive capability for the risk of HDP development. We hypothesize that additional existing data in the electronic health record, such as vital signs, medication usage, and laboratory information, would aid in improving the prediction of HDP.
To construct our machine-learning model, we first used clinical knowledge from team members to narrow the list of columns in the dataset to those relevant to hypertension. We tried several approaches to model construction including autoML approaches, evaluating performance with F1 and AUC scores. In the end, random forests with tuned parameters from a grid search yielded similar (or in many cases better) predictive performance with faster training times. To reduce racial bias in our model, we excluded race from training columns, but tracked model performance by ethnicity. In the end, our training procedure added extra copies of rows for non-Hispanic Black patients to compensate for poorer model performance on those patients, a technique called SMOTE.
Our initial model performed well, but required 200+ values to run. This would be well suited for an implementation which directly connects to an EHR, but was not realistic for routine clinical use in the form of an assessment tool. We reduced the dimensionality of the data using recursive feature elimination (RFE) to find the most important features. This yielded a smaller model, with a sufficiently small number of inputs that we could encode it into our web-based form. We are grateful to have contributed this model as a starting point to advance our understanding of hypertension in pregnancy, particularly the role of health disparities.