AI and the World's Hard Problems · Medicine & drug discovery · Nature (computational biology) · 2026-10-22
External validation of a machine learning model for predicting liver injury in children with mycoplasma pneumoniae pneumonia
Mycoplasma pneumoniae pneumonia (MPP) in children carries a high risk of liver injury, but no validated predictive tool has been available for early identification of at-risk patients. Using data from 1321 children at two centers in Nanjing plus an external validation cohort of 640, a neural network model built on six readily available predictors identified by the boruta algorithm — LDH, D-dimer, CK, age, IL-6, and CK-MB — showed the best performance. AUCs were 0.914 in training, 0.829 in testing, and 0.755 in external validation, with SHAP analysis identifying LDH as the most important predictor and an interactive web-based application developed to support early risk stratification.