<p>Explainability and clinical integration features.</p>

Ensemble methods such as Random Forest and Gradient Boosting showed strong and consistent diagnostic performance. Convolutional neural networks performed well in single-centre settings but showed reduced performance upon external validation, consistent with optimism bias. Overall, 83.6% of studies were classified as high risk of bias and 75.5% lacked external validation.

After adjusting for population size, high-income countries produced 23 times more studies per million women of reproductive age than LMICs. Sub-Saharan Africa contributed fewer than 0.1 studies per million women, and no studies were identified from Nepal or Sri Lanka. Current evidence, which is heavily concentrated in high-income and East Asian settings, is insufficient to support population-level deployment.

We introduce a four-tier LMIC feasibility classification system that links data modality, infrastructure requirements, and personnel needs. We also propose a phased research agenda that separates near-term priorities (0–2 years) from medium-term goals that depend on infrastructure investment.</p></div>