Background. Mycoplasma pneumoniae ( MP ) is a common causative pathogen of community-acquired pneumonia in children, with clinical presentations ranging in severity. Early stratification and timely intervention are essential for improving patient outcomes. However, a major clinical challenge lies in the limited ability to accurately distinguish between mild and severe cases based solely on early clinical indicators. Methods. This prospective real-world study investigated the differences in microbiome and metabolomics between mild and severe MP pneumonia (MPP) in children. Bronchoalveolar lavage fluid samples were collected from 153 children and subjected to metagenomic sequencing and non-targeted metabolomic analysis. Meanwhile, to enhance early diagnostic accuracy, this study developed a machine learning classification model and validated it using a third-party validation set. Results. The results revealed significant alterations in the abundance of specific bacterial communities in the severe group, most notably the coexistence of MP and Alphainfluenzavirus influenzae , which may contribute to disease exacerbation through synergistic pathogenic mechanisms. Furthermore, the macrolide resistant rate of MP in the severe group exceeded 80%, emphasizing the importance of appropriate antibiotic selection. Metabolomic analysis showed a significant enrichment of metabolites related to cellular energy metabolism and immune regulation in severe cases. The model demonstrated exceptional predictive performance, achieving an area under the curve ranging from 0.909 to 0.991, which significantly outperformed conventional clinical stratification methods. Conclusions. These findings elucidate the distinct pathophysiological mechanisms underlying both mild and severe MP infections and provide a promising framework for improving early diagnosis and personalized treatment strategies in paediatric MPP.