Machine Learning Improves the Predictive Utility of Lactic Acid in Hospitalized Infants

Background and ObjectivesHyperlactatemia is common in hospitalized infants. Machine learning was applied to clinical and laboratory characteristics in hospitalized infants with hyperlactatemia to identify predictors of inborn error of energy metabolism (IEEM). MethodsRetrospective cohort study of hospitalized infants aged 0-90 days (2012-2020) with a lactate [≥] 5 mmol/L. Final diagnosis was discretized to IEEM, cardiac, hypoxia, infectious and other. Random forest and XGBoost models were tuned and compared using cross-validation, and a final model was evaluated on an independent test set to determine ability to predict IEEM. ResultsAmong 1000 infants, median lactate was 8 mmol/L. The overall mortality rate was 30%, (N=291) and was 51% (N=21) among infants with an IEEM (N=41). Lactate was significantly higher in infants with an IEEM (12.6 mmol/L; IQR: 5-27 mmol/L). Machine learning analysis including plasma amino acid and acylcarnitine values yielded an area under the ROC curve (AUC-ROC) of 0.81 in a held-out test set, and was significantly better than lactate alone in a comparable population (AUC-ROC 0.81 vs. 0.56, p=0.027). ConclusionsRapid diagnosis of IEEM vs. other causes is essential for neonatal hyperlactatemia prognostication. Machine learning has high diagnostic utility, serving as a framework for computer-aided interpretation of complex diagnostic data.