Machine learning-driven risk assessment of severe Mycoplasma pneumoniae in children: analysis based on core clinical and immunological features.

OBJECTIVE

To construct and validate an interpretable machine learning model for early prediction of severe Mycoplasma pneumoniae pneumonia (SMPP) in children.

METHODS

Clinical and immunological data of 402 pediatric MPP patients were randomly divided into a training set (70% for model tuning) and a held-out internal test set (30% for final evaluation). We systematically evaluated 113 algorithms based on data collected within 24 h of admission.

RESULTS

The Gradient Boosting Machine (GBM) demonstrated optimal performance, achieving an AUC of 0.805 (95% CI: 0.724-0.886) on the held-out internal test set. The model identified 16 core predictors: dyspnea, C-reactive protein (CRP), total T lymphocytes (CD3+), sputum plug formation (PB), CD3+CD4+CD8- T cells, CD3+CD56+NKT cells, LDH, IL-6, creatine kinase (CK), PCT, ALT, IgA, abnormal coagulation function (D-dimer), CK-MB, erythrocyte sedimentation rate (ESR), and MP-DNA load. SHapley Additive exPlanations (SHAP) analysis revealed that dyspnea, elevated CRP, and decreased T lymphocytes synergistically drive severe illness risk.

CONCLUSION

The GBM-based model effectively predicts SMPP risk. Pending prospective multi-center validation, we plan to embed this transparent, data-driven tool into electronic medical record (EMR) systems to provide real-time early warning and individualized risk assessment.