Prediction of low 5-minute Apgar scores: development and internal validation of parity-stratified clinical prediction models for sub-Saharan Africa

Background Low 5-minute Apgar scores remain an important indicator of compromised neonatal condition and are associated with increased risks of morbidity and mortality. Accurate prediction of newborns at risk could guide timely intrapartum and immediate postpartum interventions. Because risk factors vary by maternal parity, prediction models may benefit from a parity-specific approach. This study aimed to develop and internally validate two prognostic models for predicting low 5-minute Apgar scores, stratified by parity. The analysis used data from 124376 singleton births at or beyond 28 weeks of gestation, recorded between July 2021 and December 2023 across 16 hospitals in Benin, Malawi, Tanzania and Uganda. Methods Model predictors were selected using a knowledge-based approach, and multivariable logistic regression was performed. Model performance was assessed through calibration and discrimination with internal validation conducted using bootstrapping. The predicted outcome was the 5-minute Apgar score, categorised as low (< 7) or normal (≥ 7). Results In the overall study population, 6.3% of newborns received a low Apgar score. The final nulliparous and parous models included 14 and 19 predictors, respectively. The models demonstrated moderate optimism-adjusted performance, with C-statistics of 0.662 for nulliparous (95% CI: 0.650 – 0.672) and 0.732 for multiparous (95% CI: 0.722 – 0.741). Calibration was excellent, with calibration-in-the-large (CITL) values of 0.000 and calibration slopes of 1.000 in both models. Antepartum haemorrhage and severe anaemia were the strongest contributors in both models. Conclusions Two prediction models for low 5-minute Apgar scores, one for nulliparous and one for parous women, demonstrated moderate predictive ability. External validation and further testing are necessary to assess the generalisability and clinical utility of these models.