Comparative Performance of Artificial Intelligence Models in Predicting Low Birth Weight: Systematic Review and Narrative Synthesis.

Abstract

Background: Low Birth Weight (LBW), defined as < 2500 g, is a global public health priority. Early prediction is essential for intervention. This systematic review investigates the performance of Artificial Intelligence (AI)/Machine Learning (ML) algorithms for LBW prediction.

Methods: Following PRISMA guidelines, Scopus, Web of Science, and PubMed were searched for studies utilizing AI/ML for LBW. Methodological quality was assessed via the PROBAST tool. Due to extreme statistical heterogeneity and inconsistent reporting of metrics, a narrative synthesis approach was employed.

Results: Forty studies met inclusion criteria. Quality assessment revealed that 82.5% were at high or unclear risk of bias, primarily due to poor reporting of calibration and data handling. Formal metaanalysis was precluded by extreme statistical heterogeneity (I2>98%) and the role of LBW as a surrogate marker for diverse phenotypes. Narrative synthesis indicated that ensemble architectures (Random Forest, XGBoost) consistently outperformed traditional linear models. Discrimination was significantly higher in high-income settings (AUC: 0.85-0.95) than in resource-limited settings (AUC: 0.65-0.75). Models integrating dynamic data, such as longitudinal ultrasound, demonstrated superior sensitivity over those relying on static clinical history. Performance trends across settings were highly variable, and given the heterogeneity, these findings must be considered strictly exploratory.

Conclusion: AI/ML models show promise for LBW prediction, but their efficacy is strictly contextdependent. Widespread methodological bias and lack of calibration metrics currently limit "bedside readiness." Clinical translation requires a transition toward population-specific, interpretable tools validated through rigorous external cohorts.