Diagnostic accuracy of artificial intelligence models for predicting hypertensive disorders of pregnancy: a systematic review and diagnostic test accuracy meta-analysis
A Prameela, H Ramteke, D Raja Kuraku, N keshapogu, R Khan, D MurugavelAbstract
Background
Hypertensive disorders of pregnancy (HDP), including preeclampsia, gestational hypertension, and HELLP syndrome, complicate 5–10% of pregnancies and are major contributors to maternal and perinatal morbidity. Beyond obstetric risk, HDP confers a 2–4-fold increased lifetime risk of cardiovascular disease. Early prediction remains suboptimal. Artificial intelligence (AI) and machine learning models integrating clinical, biochemical, and imaging data may enhance risk stratification and enable earlier preventive intervention.
Purpose
To evaluate the pooled diagnostic performance of artificial intelligence and machine learning models in predicting hypertensive disorders of pregnancy and to determine their clinical utility for early cardiovascular risk stratification.
Methods
We performed a diagnostic test accuracy meta-analysis of AI/ML models predicting hypertensive disorders of pregnancy (search through January 2026). Pooled AUC, sensitivity, specificity, and likelihood ratios were estimated using a bivariate random-effects model. Risk of bias was assessed with PROBAST-AI, and heterogeneity was explored using I² and meta-regression.
Results
A total of 12 studies comprising 164,285 pregnancies were included (mean age 30.8 years; 42.1% nulliparous). Model architectures included Random Forest (n=5), XGBoost (n=4), and Support Vector Machines (n=3). For the primary endpoint of preeclampsia prediction, AI models demonstrated a pooled AUC of 0.88 (95% CI 0.84–0.92). The pooled sensitivity was 81% (95% CI 75%–86%) and the pooled specificity was 84% (95% CI 78%–89%), with a positive likelihood ratio (LR+) of 5.06 (95% CI 3.8–6.7). For gestational hypertension, the pooled AUC was 0.82 (95% CI 0.76–0.88), with a sensitivity of 76% (95% CI 70%–82%) and a specificity of 79% (95% CI 72%–85%). HELLP syndrome prediction showed the highest diagnostic yield with a pooled sensitivity of 89% (95% CI 82%–94%) and a pooled AUC of 0.91 (95% CI 0.86–0.96), although this was based on a smaller subset of 4 studies. Subgroup analysis revealed that models utilizing multimodal data (clinical variables + uterine artery Doppler + PlGF) significantly outperformed clinical-only models (AUC 0.93 vs. 0.79; p<0.001). High heterogeneity was observed across all outcomes (I2=78.4% for sensitivity; I2=82.1% for specificity). Meta-regression identified gestational age at screening (coeff: -0.042; p=0.012) and validation method (internal vs. external; p=0.035) as significant sources of heterogeneity. Regarding the risk of bias (PROBAST-AI), 58% of studies were rated as high risk, primarily in the analysis domain (e.g., handling of missing data and over-fitting). Deeks’ funnel plot asymmetry test suggested no significant publication bias (p=0.38).
Conclusion
AI models show strong accuracy in predicting hypertensive disorders of pregnancy, particularly with multimodal inputs, supporting early risk stratification and preventive care.