Predictive models for stillbirth and neonatal death: A global scoping review of model performance, validation, transportability and implications for sub-Saharan Africa

Abstract Background: Each year 1.9 million stillbirths and 2.3 million neonatal deaths occur worldwide, and sub-Saharan Africa (SSA) accounts for 47% of stillbirths and 46% of neonatal deaths. Predictive models, including classical statistical methods, machine learning (ML) and artificial intelligence (AI), are increasingly used to detect pregnancies and neonates at high risk, but their performance and validation in SSA have not been mapped extensively. Objective: To map predictive models for stillbirths and neonatal deaths globally with a focus on SSA, and to characterise the predictors, data sources, modelling approaches, performance metrics and validation. Methods: We followed the Arksey and O'Malley framework and PRISMA-ScR guidance. We searched EMBASE, Medline, Scopus, Web of Science and Global Health from 2010 to 2025 plus reference lists. Two reviewers independently screened titles, abstracts and full texts. Data were charted across six domains: predictors, data sources, model type, evaluation metrics, performance and validation. Results: Of 1,339 records, 93 studies met the inclusion criteria; 26 (28.0%) were conducted in SSA and 67 (72.0%) elsewhere. Health facility data were the most common source (63/93, 67.7%). Classical methods alone were used in 44 studies (47.3%); ML and AI mostly appeared in hybrid combinations rather than replacing classical models. Internal validation alone was reported in 65 studies (69.9%), internal plus external in 16 (17.2%), external validation only in three (3.2%) and no validation of any kind in nine (9.7%). AUC was reported in 77 studies (82.8%), ranging from 0.610 to 0.996. Calibration-related metrics were reported in 12 studies (12.9%), four in SSA, and decision-curve analysis was reported in only one non-SSA study. Conclusion: Only one SSA study reported pragmatic clinical deployment. Most models lack external validation, calibration assessment and clinical-utility evidence for SSA. Future work should follow TRIPOD+AI and PROBAST+AI reporting standards, with pre-registered multi-site external validation in independent African cohorts.