Development of a prediction model for maternal pregnancy complications in sickle cell disease Free

Introduction Pregnant individuals with sickle cell disease (SCD) face dramatically increased maternal risks, with prior work showing over a sevenfold increase in risk of severe maternal morbidity compared to those without SCD. Despite this well-established risk, there is no validated, widely adopted tool to predict whether a pregnancy is most likely to result in maternal complications. There is an urgent need for accurate risk-stratification tools to guide preconception counseling and perinatal care.

Methods We conducted a retrospective cohort study of pregnant adults with SCD using two EMR-linked clinical data repositories: the Mass General Brigham (MGB) Research Patient Data Registry (5 affiliated hospitals) and the Penn Data and Analytics Center (3 affiliated hospitals). We identified individuals with ICD-9 or ICD-10 codes for SCD and pregnancy, and performed manual chart review to confirm SCD genotype and annotate clinical variables. We collected baseline labs (e.g., blood counts, hemoglobin quantitation), SCD treatment (e.g., hydroxyurea use, transfusion history), and complications at baseline and during pregnancy.

As a benchmark, we applied a previously published prediction model developed at Mount Sinai Hospital in Toronto (Malinowski et al., 2021). However, the Toronto model was developed using a broader composite outcome that included any transfusion or a single vaso-occlusive episode (VOE), events that may not always prompt a change in clinical management.

Given these limitations, we developed a more stringent and clinically focused composite outcome designed to better capture events with direct implications for maternal care. We defined a binary composite outcome variable reflecting severe maternal complications, which include SCD-related complications during pregnancy (e.g., acute chest syndrome [ACS], stroke, hemolytic crisis, urgent red cell exchange, sepsis, intensive care unit admission, or ≥3 VOEs requiring emergency care or hospitalization) and pregnancy-related complications (e.g., preeclampsia and venous thromboembolism). These variables were chosen based on severity, relevance, and potential to prompt changes in clinical management.

We then developed a new prediction model using multivariable logistic regression with generalized estimating equations to account for clustering by patient, trained on pregnancies not initiated on prophylactic transfusions during gestation. Predictors include genotype, hemoglobin, number of VOEs in the year prior to pregnancy, history of ACS, history of pulmonary hypertension, and an interaction term between hemoglobin level and VOE frequency. We performed internal validation using 800 bootstrap samples, a standard statistical technique to test the model's stability and assess for overfitting.

Results We identified 231 pregnancies in 167 unique patients with genotype-confirmed SCD. Most patients (n=119, 71%) contributed a single pregnancy; others contributed up to four. Genotypes included 103 HbSS, 42 HbSC, 11 HbS/β⁰-thalassemia, and 11 HbS/β⁺-thalassemia. Preventive red cell transfusions were initiated in 36 pregnancies (15.6%). A severe maternal complication (defined above in Methods) was experienced in 153 of 231 pregnancies (66.2%).

Among pregnancies not started on prophylactic transfusion (n=195), the Toronto model achieved an area under the receiver operating curve (AUC) of 0.811 using its original composite outcome. However, when we re-evaluated the Toronto model using our more stringent and clinically actionable composite outcome, its AUC declined to 0.743, indicating substantially reduced discriminative performance. Our revised model achieved an apparent AUC of 0.831 and a calibration slope of 0.908. The optimism-corrected AUC after internal validation was 0.819. These results suggest that our model achieves excellent discriminative performance with minimal overfitting.

Conclusion This study represents the largest multi-institutional effort in the United States to develop a prediction model for maternal complications in SCD. Compared to the only previously published model, our model exhibited superior discriminative performance in predicting severe maternal complications. Our model incorporates readily available clinical and hematologic predictors to risk-stratify pregnancies. Ongoing efforts will focus on refining and validating the model with the goal of creating an evidence-based tool to guide the risk-adapted management of pregnant individuals with SCD.