Objectives
The aim of this study was to develop and validate a machine learning model for distinguishing mycoplasma pneumoniae coinfection with other pathogens (Co-MPP) in children from mycoplasma pneumoniae pneumonia (MPP) in children.
Methods
Between June 2023 and March 2024, 191 consecutive pediatric patients were enrolled in this study. The latest laboratory test results before bronchoalveolar lavage (BAL) were included in the statistical analysis. After the least absolute shrinkage and selection operator (LASSO) feature screening, we input the final features into seven different machine learning classifiers (LR, SVM, KNN, Random Forest, Extra Trees, XGBoost, and LightGBM) and selected the optimal classifier for model construction. The nomogram model combined the radiomics (rad) signature and the clinical signature. The ROC curves were drawn to evaluate the diagnostic efficacy of different models. The calibration efficiency of the nomogram was evaluated by drawing calibration curves, and the Hosmer-Lemeshow test was used to evaluate the calibration ability of the models. Decision curve analysis (DCA) was utilized to evaluate the clinical utility of the models. Statistical significance was considered when the p-value was < 0.05. The statistical analysis in this study was conducted using R and SPSS 27.0 software.
Results
A total of 1834 handcrafted radiomics features were extracted, including 360 first-order features, 14 shape features, and texture features. The LR classifier achieved the best value of AUC, reaching 0.922 and 0.867 for distinguishing Co-MPP from MPP in the training and test cohorts, respectively. For building the clinical signature, LR was selected as the base model. The univariate analysis results of all clinical laboratory and CT imaging features showed that only reticulation and bronchial lumen occlusion were significantly different between MP and Co-MPP patients (p = 0.011, < 0.001, respectively). The performance showed that the clinical signature achieved AUC values of 0.729 and 0.706 in the training and test cohorts, respectively. The nomogram using the LR algorithm was performed to combine the clinical signature and rad signature. Delong test results showed the performance of the nomogram and rad signature were both higher than the clinical signature (p < 0.05), while the nomogram and rad signature showed no significant difference. Both rad signature and nomogram showed significant clinical benefit.
Conclusion
Our study demonstrated that machine learning can assist clinicians in distinguishing Co-MPP from MPP in children. Furthermore, the rad signature and nomogram model showed higher clinical benefit compared to the clinical signature.