Existing preeclampsia risk models show deteriorating performance over time

The existing prediction models for severe complications of preeclampsia are most accurate only in the two days after hospital admission, with deteriorating performance over time, according to a new study published February 4^th in the open-access journal PLOS Medicine by Henk Groen of University of Groningen, the Netherlands, and colleagues.

Preeclampsia is a potentially life-threatening condition that can occur during pregnancy; of women diagnosed with preeclampsia, 5-20% will develop severe complications. Two existing PIERS (Pre-eclampsia Integrated Estimate of RiSk) models, PIERS Machine Learning (PIERS-ML) and the logistic-regression-based fullPIERS, are designed to identify individuals at greatest or least risk of adverse maternal outcomes in the 48 hours following hospital admission for preeclampsia. However, both models are regularly used for ongoing assessment beyond the first 48 hours.

In the new study, researchers used data from 8,843 women diagnosed with preeclampsia at a median gestational age of 36 weeks between 2003 and 2016. Data included PIERS-ML and fullPIERS assessments as well as health outcomes.

The study found that neither the PIERS-ML nor fullPIERS model maintained good performance over time for repeated risk stratification in women with preeclampsia. The PIERS-ML remained generally good at identifying the very high-risk and very-low risk groups over time, but performance of the larger high-risk and low-risk groups deteriorated significantly after 48 hours. The fullPIERS model underperformed compared to the PIERS-ML model.

"Since there are no better options, clinicians may still use these two models for ongoing assessments after the first admission with pre-eclampsia, but the predictions should be treated with increasing caution as the pregnancy progresses," the authors say. More prediction models are needed that perform well over time, they add.

The authors add, "Pregnancy hypertension outcome prediction models were designed and validated for initial assessment of risks for mothers; this study shows that such 'static' models if used repeatedly over days yield increasingly inaccurate predictions."