Improved Prediction Accuracy for Late-Onset Preeclampsia Using cfRNA Profiles: A Comparative Study of Marker Selection Strategies.

Abstract

Background: Late-onset pre-eclampsia (LO-PE) remains difficult to predict because placental angiogenic markers perform poorly once maternal cardiometabolic factors dominate. Methods: We reanalyzed a publicly available cell-free RNA (cfRNA) cohort (12 EO-PE, 12 LO-PE, and 24 matched controls). After RNA-seq normalization, we derived LO-PE candidate genes using (i) differential expression and (ii) elastic-net feature selection. Predictive accuracy was assessed with nested Monte-Carlo cross-validation (10 × 70/30 outer splits; 5-fold inner grid-search for λ). Results: The best LO-PE elastic-net model achieved a mean ± SD AUROC of 0.88 ± 0.08 and F1 of 0.73 ± 0.17-substantially higher than an EO-derived baseline applied to the same samples (AUROC ≈ 0.69). Enrichment analysis highlighted immune-tolerance and metabolic pathways; three genes (HLA-G, IL17RB, and KLRC4) recurred across >50% of cross-validation repeats. Conclusions: Plasma cfRNA signatures can outperform existing EO-based screens for LO-PE and nominate biologically plausible markers of immune and metabolic dysregulation. Because the present dataset is small (n = 48) and underpowered for single-gene claims, external validation in larger, multicenter cohorts is essential before clinical translation.