Interpretable machine learning identifies a clinically inferred inflammasome-associated inflammatory injury phenotype in children with adenovirus pneumonia.

Background: Severe adenovirus pneumonia in children is characterized by persistent fever, systemic inflammation, and extrapulmonary tissue injury. Although adenovirus can activate inflammasome pathways, direct inflammasome biomarkers are not routinely available in pediatric practice. This study developed an interpretable machine-learning framework for retrospective severity discrimination and inflammatory phenotype stratification. Methods: We retrospectively analyzed 82 hospitalized children with adenovirus pneumonia, including 32 mild and 50 severe cases. Demographic, clinical, and laboratory variables were extracted from medical records. Laboratory variables were obtained from the first available blood tests within 24 hours after admission, whereas fever duration reflected the total febrile course. A surrogate Inflammasome-associated Inflammatory Injury Index (IAI) was calculated as the mean z-score of lactate dehydrogenase, C-reactive protein, aspartate aminotransferase, neutrophil-to-lymphocyte ratio, and maximal temperature. Logistic regression, support vector machine with radial basis function kernel, and random forest classifiers were evaluated using repeated stratified 5-fold cross-validation. SHAP values and unsupervised clustering were used for model interpretation and phenotype discovery. A small post-2019 cohort was used for exploratory single-center temporal validation. Results: Severe cases were younger and had longer fever duration, higher maximal temperature, lower hemoglobin, and higher alanine aminotransferase, aspartate aminotransferase, lactate dehydrogenase, and creatine kinase-MB levels than mild cases. IAI was significantly higher in severe disease. Random forest showed the best internal performance, with an AUC of 0.965 ± 0.031, PR-AUC of 0.980 ± 0.018, accuracy of 0.890, and F1-score of 0.913. Five-fold out-of-fold analysis yielded an AUC of 0.966 and a Brier score of 0.088. SHAP identified fever duration and IAI as the two leading contributors. Clustering identified an inflammatory injury-high phenotype comprising 11 patients, all with severe pneumonia. In exploratory temporal validation, the frozen random forest model achieved an AUC of 0.840 in 39 post-2019 cases. Conclusion: Routine clinical variables can support interpretable severity discrimination and phenotype discovery in pediatric adenovirus pneumonia. The inflammatory injury-high phenotype suggests a clinically detectable systemic inflammatory injury pattern biologically compatible with inflammasome-related inflammation. Because IAI is a surrogate clinical index rather than a direct molecular measure of inflammasome activation, prospective multicenter studies incorporating direct inflammasome biomarkers are required for biological validation.