Prediction of atelectasis in Mycoplasma pneumoniae pneumonia using a SHapley Additive exPlanations-interpretable...

This study aimed to evaluate the performance of three machine learning models-K-nearest neighbors (KNN), support vector machine (SVM), and neural network (NN)-for predicting the risk of atelectasis in children with Mycoplasma pneumoniae pneumonia (MPP). The study also incorporated SHAP (SHapley Additive exPlanations) analysis to interpret model predictions.

Based on the clinical data of 508 pediatric patients, we performed feature selection and developed KNN, SVM, and NN models. Model performance was compared on an independent validation set, and SHAP values were employed to elucidate the predictive logic of the models.

On the validation set, the neural network (NN) model demonstrated the best overall performance, with an AUC of 0.89 and an accuracy of 0.82. The KNN model showed comparable performance (AUC = 0.88), while the SVM model achieved the highest specificity (0.87). The SHAP analysis consistently identified neutrophil percentage (NEU.pct), serum amyloid A (SAA), and C-reactive protein (CRP) as the most critical variables influencing the predictions.

This study demonstrates the effectiveness of different machine learning models in predicting the risk of atelectasis in MPP. The neural network, in particular, exhibited superior performance owing to its powerful non-linear modeling capabilities. These interpretable models provide clinicians with a diverse set of tools to accommodate various clinical priorities, such as overall accuracy or high specificity, thereby facilitating the early identification and stratified management of high-risk children.