Abstract / Summary
Abstract Background Machine learning (ML)–based neonatal risk prediction models may support earlier identification of newborns requiring clinical attention, but predictor–outcome relationships can vary across populations, clinical settings, and outcome definitions. Understanding how candidate predictors relate to different early neonatal outcomes is, therefore, important when evaluating the contextual relevance of ML-derived risk predictor sets in routine care. Objective This study aimed to examine the associations between candidate neonatal risk predictors and three early neonatal outcomes—birth weight, bag-and-mask resuscitation, and hospitalization by Day 2—in a multisite Kenyan cohort, and to determine whether associations were consistent across outcomes. Methods We conducted a multisite observational analysis of 394 maternal-neonatal observations from three Kenyan health facilities. Candidate predictors were derived from an existing ML-based neonatal risk prediction model and included maternal, pregnancy, and delivery characteristics. Birth weight was analyzed using Spearman rank correlation, whereas bag-and-mask resuscitation and hospitalization by Day 2 were examined using bivariate Wilcoxon rank-sum tests. False discovery rate (FDR) adjustment was applied to account for multiple comparisons, with an adjusted q value of < .05 considered statistically significant. Results Seven statistically significant predictor–outcome associations involving five candidate predictors remained after FDR adjustment. Birth weight was positively associated with gestational age at delivery (ρ = 0.183; q =.002) and negatively associated with multiple birth (ρ=−0.202; q <.001) and antenatal corticosteroid exposure (ρ=−0.163; q =.006). Bag-and-mask resuscitation was associated with maternal sepsis ( q =.030) and antenatal hypertension/eclampsia ( q =.030). Hospitalization by Day 2 was associated with maternal sepsis ( q <.001) and gestational age at delivery ( q =.00018). Maternal sepsis and gestational age at delivery, therefore, demonstrated associations across 2 outcomes, whereas multiple birth, antenatal corticosteroid exposure, and antenatal hypertension/eclampsia demonstrated outcome-specific associations. Other examined predictors did not retain statistically significant associations after FDR adjustment. Conclusions Candidate neonatal risk predictors showed heterogeneous, outcome-specific associations across early neonatal outcomes in this Kenyan multisite cohort. Maternal sepsis and gestational age at delivery showed cross-outcome signals, whereas other predictors were associated with individual outcomes only. These findings provide contextual evidence regarding the empirical behavior of ML-derived candidate predictors but should not be interpreted as evidence of causality or model performance. Multivariable analyses, external validation, calibration assessment, and prospective evaluation of clinical utility are warranted before translating these associations into clinical prediction or decision support.