AI-powered risk prediction models for preventable maternal mortality in rural settings: a systematic review.

Abstract

Background: Maternal mortality remains disproportionately high in low- and middle-income countries, particularly in rural settings with limited access to skilled obstetric care. Artificial intelligence and machine learning models offer promise for early risk prediction, yet their methodological rigour, applicability, and deployment feasibility in resource-constrained rural contexts remain inadequately synthesised. This systematic review evaluated AI-powered risk prediction models for preventable maternal mortality, with an emphasis on their suitability for rural and low-resource settings.

Methods: A systematic literature search was conducted across PubMed, Scopus, Web of Science, IEEE Xplore, Google Scholar, and African Journals Online for studies published from January 2015 to August 2025. Studies employing AI or machine learning to predict maternal mortality or severe maternal outcomes were included. The Prediction model Risk Of Bias Assessment Tool (PROBAST) assessed methodological quality across four domains: participants, predictors, outcomes, and analysis. Data extraction captured study characteristics, model architectures, performance metrics, validation strategies, and rural implementation considerations. This review was registered with PROSPERO (CRD420251174343) and reported in accordance with PRISMA 2020 guidelines.

Results: Twenty-eight studies met inclusion criteria, predominantly from sub-Saharan Africa (n = 12) and South Asia (n = 8). Dataset sizes ranged from 402 to over 31 million records from national surveys (n = 14), hospital registries (n = 9), and Internet of Things monitoring systems (n = 5). Random Forest (n = 14), ensemble methods (n = 11), and neural networks (n = 11) were most frequently employed. Reported area under the receiver operating characteristic curve values ranged from 0.70 to 0.95 (median 0.84), with sensitivity 70-92% and specificity 65-85%. PROBAST assessment revealed low risk of bias for participants (24/28), predictors (25/28), and outcomes (24/28), but substantial concerns in the analysis domain (14/28 low risk, 8/28 high risk). Key limitations included reliance on synthetic oversampling without external validation, inadequate reporting of calibration, and small sample sizes in IoT studies. Only 11 studies (39%) conducted external validation. Common predictors were maternal age, blood pressure, gestational age, parity, and antenatal care attendance. Rural implementation barriers included limited connectivity, data sparsity, workforce training needs, and the absence of explainability frameworks.

Conclusions: AI-powered models demonstrate strong discrimination performance for maternal mortality prediction when trained on large, representative datasets. However, methodological weaknesses, particularly inadequate external validation and calibration assessment, limit confidence in generalizability. The underrepresentation of rural populations and the scarcity of implementation studies constrain real-world applicability. Future development should prioritise federated learning for privacy-preserving multi-site collaboration, lightweight architectures for offline deployment, explainable AI frameworks, and integration with community health worker workflows to achieve equitable, scalable solutions to reduce preventable maternal deaths in rural low- and middle-income country settings.

Systematic review registration: PROSPERO CRD42025174343.