Early Detection of Mortality Risk in Preterm Neonates Using a Random Forest Classifier and Tree-Based Pipeline Optimization Tool Automated Machine Learning

Objective: To develop and validate a machine learning–based model for early prediction of in-hospital mortality among preterm neonates using routinely available perinatal and neonatal clinical variables. Methods: This retrospective single-center cohort study included 1612 live-born preterm neonates delivered between 26 and 37 weeks of gestation. Sixteen clinical maternal, perinatal, and neonatal variables were collected, including APGAR scores (appearance, pulse, grimace, activity, respiration) , need for cardiopulmonary resuscitation (CPR), gestational age, birth weight, fetal presentation, in vitro fertilization, and obstetric history. Data preprocessing included anomaly handling, multivariate imputation by chained equations (MICE), one-hot encoding/standardization, and class imbalance management using adaptive synthetic sampling and cost-sensitive learning applied to the training dataset. Model selection compared multiple classifiers, with the final model being a Random Forest identified through the Tree-Based Pipeline Optimization Tool (TPOT) automated machine learning (AutoML) pipeline. Model performance on a held-out test set was assessed using area under the receiver operating characteristic curve (AUC-ROC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, likelihood ratios, and confusion matrix analysis. Feature importance was assessed using Chi-square/Cramér’s V, logistic regression coefficients, and model-based importance from Random Forest, gradient-boosted trees, and XGBoost. SHapley Additive exPlanations were used to determine the direction and magnitude of feature effects. Overfitting was mitigated through cost-sensitive learning, and consistency between training and test performance indicated minimal overfitting. Results: The model achieved an AUC-ROC of 0.99, accuracy of 0.98, sensitivity of 0.98, specificity of 0.99, PPV of 0.99, NPV of 0.96, and F1-score of 0.988. The most influential predictors of in-hospital mortality were need for CPR, 1-minute APGAR score, gestational age < 28 weeks, 5-minute APGAR score, birth weight < 1 kg, and fetal anomalies. Conclusion: A Random Forest–based AutoML model can accurately predict in-hospital mortality among preterm neonates using routinely available clinical variables. This data-driven tool may support earlier risk stratification, informed clinical decision-making, and improved allocation of neonatal intensive care resources in maternal-fetal and perinatal care settings.