RAMA: implementing Machine Learning to develop mortality risk prediction models for NICU patients with Acute Kidney Injury.

BACKGROUND: Acute Kidney Injury (AKI) is common in neonates admitted to the Neonatal Intensive Care Unit (NICU). Neonatal AKI is associated with multiple comorbid conditions of greater clinical severity, which also increase the neonate's risk of mortality. Understanding the risk of mortality, in addition to the severity of AKI, may be useful in determining alternative treatment options for neonates with AKI. METHODS: Two independent datasets containing neonatal patient data from eleven healthcare centers were filtered, cleaned, and combined to produce a dataset fit for training and testing with seven Machine Learning algorithms. After initial modeling with 34 training features, feature elimination was utilized to isolate the most contributing features and create streamlined models. The models were fine-tuned using Bayesian search before being compared to select the most accurate and interpretable architecture. RESULTS: A patient cohort of 245 patients included 189 alive and 56 deceased neonates with incidence of AKI. Of the 245 patients included in this study, 73.5% were male and 26.5% were female, alongside a median age at entry of 12 h and interquartile range of 35 h. Three tree-based models, Random Forest, XGBoost, and LightGBM, were found to be the most accurate and interpretable of the seven models tested. Support Vector Machine produced similar results, albeit with less interpretability. The RAMA (Raina, Arnav, Max, Aadi) model followed results of the XGBoost algorithm, with Area under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.882 ± 0.132, Accuracy of 0.878 ± 0.052, and F1 Score of 0.923 ± 0.029. CONCLUSION: RAMA utilizes a tree-based decision-making algorithm, allowing it to determine the risk of mortality in neonates susceptible to AKI.