Objective To construct the machine learning model based on routine clinical and laboratory indicators for early differentiating the Mycoplasma pneumoniae (MP), respiratory syncytial virus (RSV) and influenza A virus (FluA) infections in children with acute respiratory infections.

Methods A retrospective analysis was conducted on 1133 children diagnosed with single-pathogen acute respiratory infection (MP, RSV, or FluA) between October 2023 and December 2025. The baseline data, clinical symptoms and signs and laboratory indicators (blood routine and high-sensitivity C-reactive protein) in all cases were extracted. The patients were divided into the training group and internal validation group at a ratio of 7∶3. Additionally, 288 cases from other hospitals were included as the external validation group. LASSO regression was used to screen the predictive features, and four models, namely Logistic regression, random forest, XGBoost and LightGBM, were constructed to evaluate the discrimination, calibration and clinical net benefit, and the models were interpreted through SHAP values.

Results The XGBoost model achieved the best overall performance in internal validation, with AUC values of MP, RSV, and FluA were 0.943, 0.927, and 0.966, respectively. In external validation, the AUCs of MP, RSV and FluA were 0.754, 0.837 and 0.923, respectively. SHAP value analysis showed that the age, duration of fever, pulmonary rales, pharyngeal congestion, hypersensitive C-reactive protein, lymphocyte count and month of visit were the core predictive features.

Conclusions The discriminatory efficacy of machine learning models constructed based on conventional indicators for MP infection is significantly affected by the course of the disease and severity of the condition. It has demonstrated good generalization ability for RSV and FluA infections in external validation, and has certain application potential in primary pediatrics. It can be used as a rapid and low-cost pathogen auxiliary diagnostic tool.