A machine-learning model incorporating 8 laboratory markers demonstrated promising discrimination between severe and nonsevere influenza in hospitalized pediatric patients.
Investigators analyzed electronic medical records from 436 pediatric patients aged 28 days to 14 years who were hospitalized with influenza at Nanjing Children's Hospital in China between January 2022 and May 2023. Among them, 114 (26%) were classified as having severe influenza and 322 (74%) as having nonsevere disease. Clinical and laboratory data collected within 24 hours of admission were used to develop and evaluate 6 machine-learning algorithms. The data set was randomly divided into a training cohort of 305 patients (70%) and a testing cohort of 131 patients (30%).
The investigators aimed to develop an interpretable model for predicting severe influenza and compare its performance with alternativeapproaches. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. The investigators conducted decision curve analysis to assess potential clinical net benefit and used Shapley Additive Explanations (SHAP) to evaluate the contributions of individual laboratory variables to model predictions.
The random forest model achieved the highest AUROC at 0.88 vs. 0.87 for the support vector machine, 0.85 for LightGBM, 0.83 for the neural network model, 0.80 for AdaBoost, and 0.76 for the decision tree model. The random forest model demonstrated 84% overall accuracy, 50% sensitivity, and 97% specificity. The positive and negative predictive values were both 84%.
Decision curve analysis indicated that the random forest model provided greater net benefit than the other machine-learning algorithms across most evaluated probability thresholds. However, its sensitivity of 50% indicated that about half of severe cases in the testing cohort were not correctly identified.
To construct the prediction models, the investigators performed univariate analysis followed by least absolute shrinkage and selection operator regression to select 8 laboratory variables: neutrophil percentage, procalcitonin, aspartate aminotransferase, blood glucose, blood urea nitrogen, complement C3, B-cell percentage, and CD4-positive T-cell count.
SHAP analysis identified blood glucose as the most influential predictor, followed by aspartate aminotransferase, B-cell percentage, procalcitonin, CD4-positive T-cell count, neutrophil percentage, blood urea nitrogen, and complement C3. Higher blood glucose, aspartate aminotransferase, B-cell percentage, procalcitonin, neutrophil percentage, and blood urea nitrogen values contributed to predictions of severe influenza. Higher CD4-positive T-cell counts and complement C3 levels contributed to predictions of nonsevere disease.
Clinical comparisons revealed differences between severity groups. Bacterial coinfections were more frequent among pediatric patients with severe influenza compared with those with nonsevere disease (25% vs. 14%). The most frequently identified bacterial pathogens included Haemophilus influenzae, Staphylococcus aureus, and Streptococcus pneumoniae.
Neuraminidase inhibitor treatment prior to hospitalization was reported in 33% of patients with severe influenza vs. 48% of those with nonsevere disease. Median hospitalization duration was 13 days in the severe group vs. 6 days in the nonsevere group. All recorded mortalities occurred among patients with severe influenza, corresponding to a mortality rate of 18% within that group.
The study was limited by its retrospective, single-center design and relatively small testing cohort. The absence of external validation also leaves uncertainty about the model's performance in other pediatric populations and clinical settings. The model's 50% sensitivity indicated that a substantial proportion of severe cases were not correctly identified. Although the investigators evaluated laboratory findings collected within 24 hours of admission, the study did not establish whether the model could prospectively predict subsequent deterioration or improve clinical outcomes.
“We have developed an explainable random forest prediction model to estimate the risk of severe illness in hospitalized children with influenza,” wrote lead study author Jinhuan Wu, of the Children's Hospital of Nanjing Medical University in China, and colleagues.
The study authors declared no conflicts of interest.
