Inflammation, coagulopathy, and allergy form a “dangerous triangle” for severe HRV pneumonia in children: an interpretable machine learning study

Human rhinovirus (HRV) infection exhibits vast clinical heterogeneity in children, ranging from mild colds to severe pneumonia. Early identification of children at risk of severe progression remains a major clinical challenge, as traditional scoring systems fail to capture the complex, early-stage pathophysiological changes. This study aimed to leverage eXplainable Artificial Intelligence (XAI) to uncover the core clinical phenotype driving disease severity and develop a high-performance model for early in-hospital risk stratification. This single-center retrospective cohort study included 414 children (214 severe, 200 mild cases) hospitalized with HRV-positive pneumonia. Demographic, clinical, and laboratory data were collected. LASSO regression, applied to an initial pool of over 50 variables, selected 13 key features for model input. Six machine learning models were developed on a training set (80%) and evaluated on a hold-out test set (20%). The Random Forest (RF) model’s hyperparameters were optimized via 10-fold cross-validation and grid search. SHAP analysis identified a distinct predictive phenotype associated with severe HRV pneumonia, which we termed the “dangerous triangle”: (1) systemic inflammation (high procalcitonin), (2) coagulation-related alterations (relatively shorter APTT and higher PDW values compared with the mild group), and (3) allergic predisposition (history of allergy). The RF model, which effectively captured this phenotype, achieved excellent discrimination on the test set (AUC = 0.920), with high sensitivity (90.5%) and specificity (80.0%). Procalcitonin, platelet distribution width (PDW), and activated partial thromboplastin time (APTT) were the top three predictors. Decision curve analysis confirmed the model’s net clinical benefit across a wide range of probability thresholds (0.15–0.5). This study identified a data-driven “dangerous triangle” predictive phenotype characterized by inflammation, coagulation-related alterations, and host susceptibility in children with severe HRV pneumonia. The interpretable RF model showed strong discriminatory performance and may support early risk stratification. The observed feature associations generate hypotheses regarding the interplay among host susceptibility, inflammation, and coagulation-related responses that warrant validation in mechanistic and prospective studies.