A recent study assessed the performance of the Preeclampsia Integrated Estimate of RiSk models and the logistic regression–based model for consecutive prediction of adverse maternal outcomes in pregnant patients diagnosed with preeclampsia.
In the multi-country prospective observational study, published in PLOS Medicine, investigators analyzed data from 8,843 women admitted to maternity units across various regions between 2003 and 2016. The cohort had a median age of 31 years and a median gestational age at diagnosis of 35.79 weeks. The participants were 32% White, 30% Black, and 26% Asian.
The investigators evaluated the models' risk differentiation performance daily over a 2-week postadmission period. The primary outcome of the study was a composite of adverse maternal events, including mortality, end-organ complications (such as cardiorespiratory, renal, or hepatic issues), and uteroplacental dysfunction (eg, placental abruption).
The investigators indicated that both PIERS-ML and fullPIERS models exhibited higher mean risk predictions for the adverse outcome group compared with the uncomplicated course group. However, the discriminative capacity of both models declined over time. The area under the precision-recall curve (AUC-PRC) for the fullPIERS model remained low (range = 0.2–0.4), suggesting limited discriminative ability. The PIERS-ML model's AUC-PRC peaked at 0.65 on day 0 but decreased notably in subsequent days.
When categorizing women into multiple risk groups, the PIERS-ML model demonstrated strong rule-in capacity for the very high risk group, with positive likelihood ratios ranging from 70.99 to infinity. It also showed effective rule-out capacity for the very low risk group, where most negative likelihood ratios were 0. However, performance declined significantly for other risk categories beyond 48 hours postadmission. Decision curve analysis further revealed a diminishing advantage for treatment guided by both models over time.
"We recommend using both the PIERS-ML and fullPIERS models for consecutive prediction of adverse maternal outcome in preeclampsia more cautiously as pregnancy progresses," said lead study author Guiyou Yang, of the Department of Epidemiology at the University Medical Center Groningen at the University of Groningen in the Netherlands, and colleagues.
To enhance maternal outcomes, future research should focus on developing dynamic prediction models that account for individual patient trajectories, allowing for serial updates to risk assessments. Comparing the performance of these dynamic models with current static models will help quantify their added value in clinical practice.
The study disclosed that authors have an intellectual property claim regarding the PIERS-AI machine learning–based suite of models. No other conflicts were declared.
