Diagnostic Machine Learning Models of Infectious Mononucleosis in Children Based on Clinical Data: A Retrospective Multicenter Study

ABSTRACT The clinical manifestations of infectious mononucleosis (IM) and acute respiratory tract infections (ARTI) exhibit significant similarities. We aim to develop cost‐efficient models for IM in children utilizing the Shapley Additive explanation (SHAP) algorithm. We conducted a retrospective analysis of patients with the first diagnosis of IM from three medical centers. We employed four different machine learning techniques to develop new diagnostic models based on clinical features and serum inflammatory markers. The predictive accuracy of model was evaluated using the ROC curve and compared with traditional indicators. This study included a total of 853 patients with 49 clinical features. Through ten‐fold cross‐validation, the best‐performing integrated learning models are GBM, XGBoost, and RSF. The models were interpreted using SHAP to derive the feature subsets Lymphocyte, PLR, LDH, SII, Age, these subsets comprised the final diagnostic prediction model. The results show that the models based on five indicators have the same IM diagnostic performance as the EBV‐specific examination, and have a higher diagnostic value than the diagnosis based on atypical lymphocytes and EBV‐DNA load. Meanwhile, our models are applicable to children with IM of different age groups. This study provides a new diagnostic tool for differentiating IM from ARTI in children. Our novel diagnostic models are independent of EBV‐specific test results and exhibit superior diagnostic performance compared to traditional markers in the diagnosis of IM, particularly for primary healthcare units and institutions lacking EBV‐specific detection capabilities.