Development and validation of a machine learning model for predicting co-infection of Mycoplasma pneumonia in pediatric patients.

Abstract

Background: Mycoplasma pneumoniae pneumonia (MPP) is endemic in China, while Mycoplasma co-infection with other pathogens (Co-MPP) linked to severe outcomes. Despite radiomics and machine learning potential in pneumonia, pediatric Co-MPP differentiation remains underexplored. This study aimed to bridge this gap by evaluating machine learning models, particularly radiomics features derived from high-resolution computed tomography (HRCT) scans, to differentiate between MPP and Co-MPP, and to compare their predictive performance with traditional clinical models.

Methods: We conducted a retrospective analysis of hospitalized pediatric pneumonia patients from June to December 2023 at Affiliated Hospital of Xuzhou Medical University. Chest computed tomography (CT) scans were performed using a multi-slice CT scanner with over 64 detectors. Fluorescent quantitative polymerase chain reaction (PCR) was used to detect 14 pathogens in bronchoalveolar lavage (BAL) fluid. The most recent laboratory results prior to BAL were included in multifactorial logistic regression (LR) analysis, selecting variables with P<0.05 for constructing the clinical model. The largest cross-section of the lesion was selected, and image segmentation was performed using ITK-SNAP software. Radiomics features were extracted with Pyradomics. Features were filtered using t-tests, Mann-Whitney U tests, and Spearman rank correlation coefficients. The least absolute shrinkage and selection operator (LASSO) regression and ten-fold cross-validation were used for feature selection and to construct the radiomics model, optimizing the dimensionality of the dataset. Eight different machine learning models [LR, support vector machine (SVM), K-nearest neighbor (KNN), RandomForest, ExtraTrees, eXtreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and multi-layer perceptron (MLP)] were trained with the selected features, with five-fold cross-validation yielding the final radiomics model. The clinical and radiomics models were combined to create a nomogram model. Data analysis was performed using R software and SPSS 26.0.

Results: A total of 124 cases of MPP and children with Co-MPP were included. The extracted radiomics features consisted of first-order signal intensity features (n=360), morphological features (n=14), and texture features (n=1,460). LASSO regression and ten-fold cross-validation identified 23 non-zero correlation coefficient features for constructing Radscore. The LR model demonstrated superior predictive performance for Co-MPP in the validation cohort, with an area under the curve (AUC) of 0.951, sensitivity of 0.778, and specificity of 0.875. The nomogram model combining clinical and radiomics labels significantly outperformed the clinical model (P=0.004). Calibration curve analysis indicated that the nomogram model exhibited the best agreement with actual values. Both the radiomics and nomogram models provided greater clinical net benefits compared to the clinical model.

Conclusions: The radiomics model trained using machine learning effectively predicts Co-MPP in children, while the combined clinical and radiomics nomogram model offers the best predictive performance.