Predictive Model for Preeclampsia Combining sFlt-1, PlGF, NT-proBNP, and Uric Acid as Biomarkers

Preeclampsia (PE) is a hypertensive, multisystem disorder affecting 2–5% of pregnancies and is a leading cause of maternal and perinatal morbidity. Pathophysiology involves placental dysfunction with imbalance of angiogenic factors, notably elevated sFlt-1 and reduced PlGF; the sFlt-1/PlGF ratio is used clinically to aid diagnosis and prognosis, with a ratio <38 between 24+0 and 36+6 weeks effectively ruling out PE for up to 4 weeks. Cardiovascular strain in pregnancy is abnormal in PE, and cardiac biomarker NT-proBNP and serum uric acid levels are higher in PE and correlate with severity. Individually, NT-proBNP and uric acid have limited predictive performance, but combined models may improve prediction. The study’s objectives were to externally validate a previously published machine-learning model combining sFlt-1, PlGF, NT-proBNP, and uric acid to predict preterm PE in women with clinical suspicion, and to compare its performance with the sFlt-1/PlGF ratio alone.

The sFlt-1/PlGF ratio is a validated predictor for PE and adverse outcomes, with established rule-out thresholds and known associations with time to delivery. NT-proBNP is elevated in hypertensive disorders of pregnancy and correlates with PE severity and earlier onset, although as a single marker it provides modest prediction. Uric acid is commonly elevated in PE and associates with worse outcomes, but its prognostic value is debated. Prior work suggested that adding NT-proBNP (and uric acid) to the sFlt-1/PlGF ratio could improve positive predictive value and specificity for early-onset PE and for short-term prediction of delivery, warranting external validation of such combined models.

Design: External validation using a real-world, multicenter cohort of pregnant women with suspected PE between 24+0 and 36+6 weeks, recruited from seven Spanish University Hospitals (EuroPE study cohort) from March 2018 to December 2020. Multiple samples per patient were allowed (restricted to one per gestational week). Ethics approval obtained and written informed consent provided. Exclusions: gestational age outside 24+0–36+6, multiple gestation, fetal anomalies, loss to follow-up, and conditions requiring immediate delivery. Diagnostic criteria: ISSHP criteria for PE (BP >140/90 mmHg on two occasions ≥4 h apart plus proteinuria >300 mg/24 h or dipstick ≥2+ after 20 weeks). Early-preterm PE <34+0 weeks; late-preterm PE 34+0–36+6 weeks. Suspicion of PE based on hypertension/proteinuria changes, symptoms (e.g., headache, visual disturbance, epigastric pain), abnormal labs (thrombocytopenia, elevated liver enzymes), abnormal uterine artery Dopplers, or IUGR by local standards. Outcomes: Adverse maternal (ICU admission, eclampsia, placental abruption, DIC, pulmonary edema, HELLP) and perinatal/neonatal outcomes as predefined. Laboratory methods: Serum PlGF, sFlt-1, NT-proBNP measured on Roche Cobas e601 (electrochemiluminescence); uric acid measured on Abbott Alinity c (uricase method). Turnaround times: PlGF and sFlt-1 18 min; NT-proBNP 9 min; uric acid 10 min. Measuring ranges and limits of quantification noted; intra/inter-assay CVs <5%. Statistical analysis: Demographics with SPSS 26; normality via Kolmogorov-Smirnov; ANOVA/Chi-squared as appropriate with p<0.05. The machine-learning model (random forest) included six predictors: gestational age at admission, chronic hypertension, and GA-corrected serum levels of sFlt-1, PlGF, NT-proBNP, and uric acid. Repeated measurements per patient were categorized into risk levels (low to high). A patient was defined negative if none of the repeated measurements reached moderately high or above; otherwise positive. Performance metrics (sensitivity, specificity, PPV, NPV, AUC) with 95% CIs via bootstrap (10,000 samples). Implementation with scikit-learn 0.23.2. Decision thresholds are provided in supplementary materials.

Cohort: 597 women (936 serum samples) between 24+0 and 36+6 weeks after exclusions; overall PE incidence 34.7% (207/597): early-preterm PE 15.1% (90), late-preterm PE 11.2% (67), term PE 6.3% (50), and no PE 65.3% (390). Women who developed PE had significantly higher sFlt-1, sFlt-1/PlGF ratio, NT-proBNP, and uric acid than those without PE (p<0.001), with highest levels in early-preterm PE. Early-preterm PE had earlier delivery, higher cesarean rates, more maternal OICU admissions, lower birthweight, higher IUGR, worse Apgar, more NICU admissions, and more adverse neonatal outcomes. Overall preterm PE prediction (Table 2): Compared with sFlt-1/PlGF alone, the combined model showed improved specificity and PPV and similar or slightly better sensitivity/NPV:

• Sensitivity: 79.6% vs 77.5% (p=0.210)
• Specificity: 94.9% vs 91.0% (p<0.05)
• PPV: 83.1% vs 72.8% (p<0.05)
• NPV: 93.7% vs 92.8% (p=0.140) False-positive and false-negative rates were reduced with the model vs sFlt-1/PlGF alone (FPR 23% vs 41%; FNR 29% vs 32%). Early-preterm PE (Table 3):
• Sensitivity: 86.7% vs 82.2%
• Specificity: 93.8% vs 90.8% (improved)
• PPV: 80.4% vs 72.5% (improved)
• NPV: 96.0% vs 94.5% Late-preterm PE (Table 3):
• Sensitivity: 63.5% vs 63.5%
• Specificity: 96.0% vs 90.2% (improved)
• PPV: 75.0% vs 55.0% (improved)
• NPV: 93.3% vs 92.9% Short-term prediction (Table 4): Within 1 week, early-preterm PE PPV and specificity improved (PPV 78.7% vs 70.5%; specificity 93.8% vs 90.8%; sensitivity 90.9% vs 87.0%; NPV 97.6% vs 96.5%). For late-preterm PE within 1 week, PPV and specificity improved (PPV 74.4% vs 54.2%; specificity 96.0% vs 90.2%) with unchanged sensitivity (65.3%). Within 3 weeks (early-preterm) or 2 weeks (late-preterm), specificity and PPV also tended to be higher with the combined model, with variable sensitivity. Overall, the combined model outperformed sFlt-1/PlGF ratio alone, particularly by increasing PPV and specificity and reducing false positives, with advantages more evident as delivery approached and in early-preterm PE.

By externally validating a machine-learning model that integrates placental (sFlt-1/PlGF), cardiac (NT-proBNP), and renal (uric acid) biomarkers, the study shows improved ability to rule in and, slightly, to rule out preterm PE compared with the sFlt-1/PlGF ratio alone. The enhanced PPV and specificity, especially for early-preterm PE and for short-term horizons (within 1–3 weeks), suggest fewer false positives and more precise identification of women who will develop PE imminently. This addresses a clinical gap wherein high sFlt-1/PlGF ratios can yield many false positives requiring intensive monitoring. Integrating NT-proBNP likely captures maternal cardiovascular maladaptation, and uric acid reflects renal involvement, complementing angiogenic imbalance and leading to better discrimination. These findings align with prior reports that combining NT-proBNP with sFlt-1/PlGF improves prediction and reduces false positives, and support incorporating multi-system biomarkers to refine PE risk stratification.

A machine-learning model combining the sFlt-1/PlGF ratio, NT-proBNP, and uric acid predicts preterm PE better than the sFlt-1/PlGF ratio alone, notably increasing specificity and PPV and reducing false positives, with improved short-term prediction (within 1 and 3 weeks). Implementation could reduce unnecessary interventions and optimize monitoring in suspected PE. Future work should test clinical utility, cost-effectiveness, and generalizability, refine cutoffs across platforms, and explore integration into decision-support software and prediction of maternal-fetal adverse outcomes.

The study is an external validation within a nested cohort not managed based on biomarker results; although rigorous, randomized trials are needed to determine clinical impact on hospitalization and costs. The sFlt-1/PlGF cutoffs are assay- and platform-specific (Elecsys), limiting generalizability to other platforms. The analysis could not assess 4-week predictions in some subanalyses. Author affiliations and detailed mapping are not fully provided here. Generalizability to broader populations and multiple gestations is limited by exclusions.