Explainable machine learning for diagnosing severe Mycoplasma pneumoniae pneumonia in children: model development and internal validation.

BACKGROUND

Mycoplasma pneumoniae pneumonia (MPP) is common in children, but severe MPP (SMPP) may progress rapidly and is difficult to distinguish from non-severe disease because clinical, laboratory, and radiographic findings overlap. We aimed to develop and internally validate machine-learning models for adjunctive SMPP risk stratification using routine data while minimizing circular reasoning.

METHODS

We retrospectively included consecutive children with MPP admitted from January 1, 2015, to January 1, 2026. Demographic, clinical, laboratory, and radiographic variables were collected. Variables overlapping with the severity definition, severity-proximal biomarkers, and model-derived leakage variables were excluded before modeling. Eight supervised algorithms were developed in a training cohort and evaluated in an internal test cohort using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1-score, calibration, Brier score, and decision curve analysis.

RESULTS

The cohort included 1,046 children, of whom 205 had SMPP; 784 were assigned to the training set and 262 to the internal test set, including 51 SMPP cases. In the strict non-overlap test set, the support vector machine achieved the highest AUC of 0.947 [95% confidence interval (CI), 0.907-0.980], with accuracy of 0.924, sensitivity of 0.784, specificity of 0.957, F1-score of 0.800, and Brier score of 0.058. Random forest achieved an AUC of 0.926 (95% CI, 0.878-0.964), accuracy of 0.897, sensitivity of 0.824, specificity of 0.915, and Brier score of 0.094. It was retained for calibration, decision-curve, threshold, and feature-importance analyses because of its interpretability and balanced performance. Important predictors included aspartate aminotransferase, alanine aminotransferase, albumin, cough duration, blood urea nitrogen, white blood cell count, platelet count, age, wheezing, and lung rales.

CONCLUSION

After exclusion of leakage variables and severity-definition-overlapping predictors, machine-learning models maintained good internal performance for classifying SMPP in children with MPP. They should be considered adjunctive risk-stratification tools rather than standalone early diagnostic tools. Multicenter external validation and prospective workflow evaluation are required before clinical implementation.