Artificial intelligence-powered models in predicting mortality in maternal, newborn, and children under five: a...

Maternal, newborn, and under-five mortality remain significant global health challenges, with 287,000 maternal deaths reported in 2020 and 4.9 million under-five deaths in 2022. The use of artificial intelligence in prediction models has shown great potential in predicting these mortalities. Artificial intelligence (AI)-powered models are capable of processing complex datasets to identify non-linear relationships and patterns, potentially outperforming traditional statistical methods in predictive accuracy, thereby enabling early detections and interventions more specifically. As AI develops rapidly, it is imperative to carry out systematic evaluations and comparisons of the accuracy, sensitivity, and applicability of these models to ensure the most efficient and effective use of AI in reducing mortality in these groups. The objective of this protocol is to outline the methods that will appraise, synthesise, and compare the existing literature in this domain.

The review will follow the Preferred Reporting Items for Systematic Reviews and Meta-Analysis 2020 guidelines, implementing a comprehensive search strategy across 14 databases, including PubMed, Embase, and Scopus. The review will refer to the most recent multiple methodological guidelines applicable to this study to determine inclusion and exclusion criteria and the content of the included literature. The articles will then be integrated, assessed, and compared for transparency, standardisation, and fairness, as well as model performance, generalisation, and heterogeneity, in accordance with the aforementioned methodological guidelines. Where possible, meta-analyses of the performance of models meeting specific criteria will also be conducted to better identify more effective models.

This protocol outlines the methodology used in a systematic review of artificial intelligence prediction models for predicting maternal, neonatal, and under-five mortality. Key anticipated challenges include comprehensively identifying bias and heterogeneity, establishing reasonable and transparent standards for quantitative synthesis, and addressing statistical dependence introduced multiple models within original studies. To address these, we will apply a structured taxonomy of bias, pre-defined tiered synthesis rules, and methods such as robust variance estimation and bivariate models. We expect that this review will fill important gaps in the field and provide insights that support the advancement of artificial intelligence in maternal and child healthcare.

PROSPERO CRD42024569282.