Efficacy and comparative performance of machine learning models for stroke risk prediction in hypertensive patients:...

Stroke poses a significant health burden among hypertensive patients, where traditional risk models often lack precision. Machine learning (ML) has shown promise in enhancing prediction accuracy by integrating diverse data sources.

Following PRISMA guidelines, we searched 5 databases from inception to September 2025. Eligible studies reported the performance of ML models in hypertensive cohorts. Data were pooled using random-effects models, with heterogeneity assessed via I², subgroup analyses, meta-regression, and leave-one-out sensitivity. The risk of bias was evaluated using PROBAST + AI, and the evidence quality was assessed using the GRADE approach.

Ten studies (n = 13,299 stroke cases) were included. Pooled sensitivity was 0.88 (95 % CI: 0.80-0.93), specificity 0.88 (95 % CI: 0.77-0.94), positive likelihood ratio 7.1 (95 % CI: 3.4-15.1), negative likelihood ratio 0.14 (95 % CI: 0.08-0.26), and AUC-ROC 0.94 (95 % CI: 0.91-0.96), indicating good discriminative ability. Heterogeneity was high for both sensitivity (I² = 79.5 %) and specificity (I² = 76.8 %), potentially due to variations in study design and populations. Subgroup analyses showed consistent performance in Chinese studies (sensitivity 0.85, specificity 0.84) and those using multimodal features (sensitivity 0.84, specificity 0.83), with higher sensitivity for ischemic/hemorrhagic-specific models (0.90). Meta-regression explained 73.9 % of variance and No publication bias was detected (Deeks' p = 0.654).

ML models demonstrate good performance for stroke prediction in hypertensive patients. However, heterogeneity underscores the need for standardized approaches. This evidence, rated moderate by GRADE, supports ML integration in clinical practice for improved prevention.