Study finds accuracy of preeclampsia risk models decline over time

Study finds accuracy of preeclampsia risk models decline over time | Image Credit: © vchalup - © vchalup - stock.adobe.com.

The accuracy of current models for predicting severe complications of preeclampsia is highest only at day 2 after hospital admission and decreases over time, according to a recent study published in PLOS Medicine.¹

Severe complications of preeclampsia occur in 5% to 20% of women with the condition. The risks of these severe maternal outcomes are often determined using 2 Pre-eclampsia Integrated Estimate of RiSk (PIERS) models: PIERS Machine Learning (PIERS-ML) and the logistic-regression-based fullPIERS.

While these models were designed for use within the 48 hours after hospital admission, they are often used for ongoing assessment after this timeframe. According to investigators, “neither PIERS-ML nor fullPIERS has been validated for such application.”²

The prospective observational study was conducted to determine the efficacy of these models for ongoing prediction of adverse maternal outcomes in patients with preeclampsia. A database of 8843 women with a median gestational age of 35.79 weeks at preeclampsia diagnosis was included in the analysis.

Maternity units from the Americas, sub-Saharan Africa, South Asia, Europe, and Oceania were assessed, with data obtained from 2003 to 2016. Preeclampsia was determined based on 2021 International Society for the Study of Hypertension in Pregnancy criteria. Patients underwent follow-up at routine prenatal and postnatal clinical visits.

Having at least 1 adverse maternal outcome was considered the primary measure of the analysis. Adverse outcomes included maternal mortality, severe maternal morbidity impacting certain maternal systems, and other serious complications such as placental abruption.

Women with an uncomplicated course were those without any of these outcomes. Predictor measurements for the 2 models and outcome assessments were conducted using each site’s clinical guideline.

Of participants, 32% were White, 30% Black, and 26% Asian. Additionally, 38.1% were recruited from Canada, with a mean follow-up time of 5 days reported. An adverse outcome was reported in 1083 of 8843 women.

Women were aged a median 31 years at the estimated due date and were at a median gestational age of 35.79 weeks at admission. The number of patients with measurements consistently decreased from day 0 to day 13 after admission.

On day 0, the peak mean predicted probability of adverse outcomes in the next 48 hours among patients with adverse outcomes was reported for the PIERS-ML model. After this point, a steady decline was observed until day 3, after which the probability shifted between 0.1 and 0.2.

For patients with uncomplicated course, the PIERS-ML model had the lowest mean predicted probability of adverse outcomes at day 0, increasing until a peak was reached on day 2 and remained stable. The mean predicted probability remained significantly higher in the adverse outcomes group vs the uncomplicated course group.

Similar patterns were observed when using the fullPIERS model. The adverse outcome group also had significantly higher values with fluctuations over time based on this model, while lower, more stable values were reported in the uncomplicated course group.

Area under the precision-recall curves (AUC-PRC) for the PIERS-ML model were 0.65 on day 0 and 0.52 on day 1. After this point, AUC-PCR values ranged from 0.1 to 0.5. In comparison, AUC-PCR values on the fullPIERS model ranged from 0.2 to 0.4 on most days.

These results indicated significantly accuracy of the models for predicting preeclampsia at only day 2 after admission. Investigators recommended “using both the PIERS-ML and fullPIERS models for consecutive prediction of adverse maternal outcome in preeclampsia more cautiously as pregnancy progresses.”

References

1. Study probes how to predict complications from preeclampsia. PLOS. February 4, 2025. Accessed February 10, 2025. https://www.eurekalert.org/news-releases/1071787
2. Yang G, Montgomery-Csobán T, Ganzevoort W, et al. Consecutive prediction of adverse maternal outcomes of preeclampsia, using the PIERS-ML and fullPIERS models: A multicountry prospective observational study. PLOS Medicine. 2025. doi:10.1371/journal.pmed.1004509