Objective:

To develop an interpretable machine-learning model for predicting severe influenza in hospitalized children and compare its performance with other algorithms.

Approach:

• Study population: Retrospective electronic medical records for 436 children aged 28 days to 14 years hospitalized with influenza at Nanjing Children's Hospital, China, from January 2022 to May 2023. Severe influenza was recorded in 114 patients (26%).
• Model development: Admission clinical and laboratory data collected within 24 hours were used to train and test six algorithms. The dataset was randomly split into training (305 patients) and testing (131 patients) cohorts.
• Markers and evaluation: Eight selected laboratory variables were neutrophil percentage, procalcitonin, aspartate aminotransferase, blood glucose, blood urea nitrogen, complement C3, B-cell percentage, and CD4-positive T-cell count. Performance was assessed using discrimination and cl…

Key Findings:

• The random forest model had the highest AUROC (0.88), compared with 0.87 for support vector machine, 0.85 for LightGBM, 0.83 for neural network, 0.80 for AdaBoost, and 0.76 for decision tree.
• In the testing cohort, random forest accuracy was 84%, sensitivity 50%, specificity 97%, and positive and negative predictive values were each 84%.
• Decision curve analysis indicated greater net benefit for random forest than the other algorithms across most evaluated probability thresholds.
• SHAP ranked 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 glucose, aspartate aminotransferase, B-cell percentage, procalcitonin, neutrophil percentage, and blood urea nitrogen contributed to severe-disease predictions; higher CD4-positive T-cell counts and complement C3 contributed to nonsevere-disease predict…
• Bacterial coinfections were more frequent in severe than nonsevere cases (25% vs. 14%). Median hospitalization was 13 versus 6 days, respectively; all recorded deaths occurred in the severe group, where mortality was 18%.

Interpretation:

The authors reported developing an explainable random forest model to estimate severe-illness risk in hospitalized children with influenza. Its 50% sensitivity meant that about half of severe cases in the testing cohort were not correctly identified.

Limitations:

• Retrospective, single-center design and relatively small testing cohort.
• No external validation, leaving performance in other pediatric populations and settings uncertain.
• The study did not establish whether the model prospectively predicts later deterioration or improves clinical outcomes.

Conclusion:

The authors reported developing an explainable random forest model to estimate the risk of severe illness in hospitalized children with influenza. They declared no conflicts of interest.

Sources:

• BMC Medical Informatics and Decision Making (as cited in the provided context)