Machine Learning-Based Prediction of Adverse Neonatal Outcomes in Preterm Births Across Nigerian Neonatal Units: A Multi-Centre Prospective Study

Adverse neonatal outcomes in preterm births, including respiratory distress syndrome, necrotising enterocolitis, intraventricular haemorrhage, and neonatal death, impose a severe burden on Nigerian neonatal intensive care units, yet validated clinical prediction tools calibrated to Nigerian neonatal unit conditions are absent. This dissertation develops and validates machine learning models for predicting adverse neonatal outcomes in preterm births across Nigerian neonatal units. A prospective multi-centre cohort enrolled 480 preterm births at neonatal units in Lagos, Kano, Enugu, Bayelsa, Plateau, and Adamawa States. Clinical predictor variables including gestational age, birth weight, delivery mode, antenatal corticosteroid administration, Apgar scores, admission temperature, and early laboratory parameters were collected. Five machine learning algorithms were trained and compared: logistic regression, random forest, gradient boosting, support vector machine, and a deep learning multilayer perceptron. The gradient boosting model achieved the highest AUC of 0.91 for composite adverse outcome prediction, significantly outperforming the logistic regression baseline (AUC 0.79). SHAP value analysis identified admission hypothermia, gestational age below 28 weeks, and absence of antenatal corticosteroids as the three highest-contribution predictors. External validation at a seventh site achieved AUC of 0.88. A clinically deployable risk stratification tool was developed. The dissertation provides an original machine learning prediction framework for Nigerian preterm neonatal care and advocates for its integration into admission triage protocols at resource-limited neonatal units. Keywords: preterm birth, neonatal outcome prediction, machine learning, Nigeria, neonatal intensive care