Clinical Key Features Uncovered by Blood Eosinophilia‐Based Machine Learning Classification of Chronic Rhinosinusitis

Background Classification based on the existence and severity of blood eosinophilia via machine learning (ML) may provide novel insights into the pathophysiology of chronic rhinosinusitis (CRS). Methods Key features of CRS with blood eosinophilia were investigated through exploratory data analyses and ML, focusing on the existence (Setting‐1: absolute eosinophil count [AEC] cutoff: 500/µL) and severity (Setting‐2: AEC cutoffs: 500 and 1500/µL) of blood eosinophilia. Four ML models were tested for each setting; SHapley Additive exPlanations (SHAP) was applied to identify key classification features of the best model. Results Univariate analyses targeting 399 patients with CRS demonstrated significant differences for 17 and eight additional features in both settings and Setting‐2, respectively. Setting‐2 revealed an increased incidence of eosinophilic CRS without nasal polyps (NPs) with increasing severity. Random forest and eXtreme Gradient Boosting were the best ML models in Setting‐1 and ‐2, respectively. Based on SHAP, the blood basophil count was one of the three most important classification features for any class. In Setting‐2, high C‐reactive protein levels, high blood basophil count, comorbid chronic eosinophilic pneumonia, low computed tomography (CT) scores in the maxillary sinus, and low NP scores were the five most important classification features separating an AEC ≥ 1500/µL from other classes. Conclusion ML classification of CRS revealed the involvement of basophils in eosinophilia in peripheral blood. Features specific for CRS with an AEC ≥ 1500/µL indicated that otolaryngologists should suspect the potential eosinophilic CRS with unique comorbidities even in the absence of NPs and the unremarkable inflammation on sinonasal CT.