Objective:

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

Approach:

• Data Collection: Clinical and immunological data of 402 pediatric MPP patients were collected and divided into a training set (70%) and a held-out internal test set (30%).
• Model Evaluation: 113 algorithms were systematically evaluated based on data collected within 24 hours of admission.
• Model Selection: The Gradient Boosting Machine (GBM) was identified as the optimal model, achieving an AUC of 0.805 on the test set.
• Predictor Identification: The model identified 16 core predictors influencing the risk of severe illness.
• SHAP Analysis: SHapley Additive exPlanations (SHAP) analysis was used to reveal key drivers of severe illness risk.

Key Findings:

• The GBM model achieved an AUC of 0.805 (95% CI: 0.724–0.886) for predicting SMPP.
• 16 core predictors were identified, including dyspnea, CRP levels, total T lymphocytes, sputum plug formation, CD3+CD4+CD8− T cells, CD3+CD56+NKT cells, LDH, IL-6, creatine kinase, PCT, ALT, IgA, abnormal coagulation function, CK-MB, erythrocyte sedimentation rate, and MP-DNA load.
• SHAP analysis indicated that dyspnea, elevated CRP, and decreased T lymphocytes synergistically increase the risk of severe illness.

Interpretation:

The model highlights the importance of integrating clinical and immunological data.

Limitations:

• The model requires prospective multi-center validation.
• Current research lacks systematic integration of specific immunophenotypes and comprehensive immune features.

Conclusion:

The study presents a transparent, data-driven tool for early warning and individualized risk assessment in pediatric SMPP.