METHODS: This national-level, multicentre, prospective direct observational study was conducted in neonatal intensive care units (NICUs) of five public hospitals in Malaysia. Randomly selected nurses were directly observed during medication preparation and administration. Each observation was independently assessed for errors. Ten machine learning (ML) algorithms were applied with features derived from systematic reviews, incident reports, and expert consensus. Model performance, prioritising F1-score for MAEs, was evaluated using various measures. Feature importance was determined using the permutation-feature importance for robust comparison across ML algorithms.
RESULTS: A total of 1093 doses were administered to 170 neonates, with mean age and birth weight of 33.43 (SD ± 5.13) weeks and 1.94 (SD ± 0.95) kg, respectively. F1-scores for the ten models ranged from 76.15% to 83.28%. Adaptive boosting (AdaBoost) emerged as the best-performing model (F1-score: 83.28%, accuracy: 77.63%, area under the receiver operating characteristic: 82.95%, precision: 84.72%, sensitivity: 81.88% and negative predictive value: 64.00%). The most influential features in AdaBoost were the intravenous route of administration, working hours, and nursing experience.
CONCLUSIONS: This study developed and validated an ML-based model to predict the presence of MAEs among neonates in NICUs. AdaBoost was identified as the best-performing algorithm. Utilising the model's predictions, healthcare providers can potentially reduce MAE occurrence through timely interventions.