Incremental value of biophysical and biochemical first trimester markers for predicting early and late preeclampsia: a nested model analysis
Aim : to compare the discriminative ability and calibration of nested-architecture machine learning (ML) models and the Fetal Medicine Foundation (FMF) algorithm for predicting early-onset (< 34 weeks) and late-onset (≥ 34 weeks) preeclampsia (PE) in the first trimester of pregnancy in Russian population. Materials and Methods. with 23,247 singleton pregnancies was carried out. The biophysical sample comprised 22,230 observations, and the biochemical subsample included 7,581 observations. Nested ML models were developed: M0 (maternal characteristics) → M2 [(M0 + biophysical markers – mean arterial pressure (MAP) and uterine artery pulsatility index (UtAPI)] → M4 (M2 + biochemical markers – placental growth factor (PlGF), soluble fms-like tyrosine kinase-1 (sFlt-1) and pregnancy-associated plasma protein-A (PAPP-A)]. A comparison was made with the FMF algorithms – prior FMF risk (FMF prior) and posterior FMF risk (FMF posterior). Metrics included: AUC-ROC (Area Under the Receiver Operating Characteristic curve) with 95 % confidence interval (CI) calculated by the bootstrap method (n = 2000); the observed-to-expected ratio (O:E ratio); incremental value of biomarkers – change in C-statistic (ΔC), net reclassification improvement (NRI), integrated discrimination improvement (IDI), and sensitivity at fixed false-positive rates. Statistical significance of AUC differences was assessed using the DeLong test. Results. The AUC of model M4 for early-onset PE (EOPE) was 0.972 (95 % CI = 0.947–0.991) compared with 0.951 for FMF posterior (p = 0.368); for late-onset PE (LOPE), the AUC was 0.871 (95 % CI = 0.833–0.906) versus 0.822 for FMF posterior (p = 0.009). After Platt scaling recalibration, the ML models demonstrated adequate calibration (O:E ≈ 1.0). FMF posterior demonstrates inadequate calibration for LOPE in the Russian population (O:E = 3.34; 95 % CI = 2.61–4.21), indicating 3.34-fold more LOPE cases than predicted by the model. Biophysical markers provided 4-fold greater incremental value for EOPE (ΔC = +0.029 vs. +0.007), while biochemical markers were more valuable for LOPE (ΔC = +0.020 vs. +0.001). For EOPE the UtAPI was a lead predictor (odds ratio (OR) = 3.28; 95 % CI = 2.68–4.70), for LOPE – previous PE (OR = 1.90; 95 % CI = 1.72–2.09). Conclusion . ML models significantly outperform the FMF algorithm in discriminating late preeclampsia and provide substantially better calibration in Russian population. The inadequacy of the FMF algorithm for LOPE (O:E = 3.34 indicates 3.3-fold more LOPE cases than predicted) justifies the need for specialized models for this phenotype. Differential incremental value of biomarkers was established: biophysical markers (UtAPI) are critical for EOPE, biochemical markers (PlGF, sFlt-1) – for LOPE, supporting the concept of contingent screening.