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

Area of Science:

• Medical Research
• Data Science in Medicine
• Otolaryngology

Background:

• Chronic rhinosinusitis (CRS) pathophysiology can be illuminated by classifying disease based on blood eosinophilia using machine learning (ML).
• Investigating the existence and severity of blood eosinophilia in CRS patients offers novel insights.

Purpose of the Study:

• To explore key features of CRS with blood eosinophilia using ML.
• To classify CRS based on absolute eosinophil count (AEC) cutoffs of 500/µL (Setting-1) and 500/1500/µL (Setting-2).

Main Methods:

• Exploratory data analyses and four ML models were employed for classification in two settings.
• SHapley Additive exPlanations (SHAP) identified key classification features for the best-performing models.
• Univariate analyses included 399 CRS patients.

Main Results:

• ML models, specifically random forest (Setting-1) and eXtreme Gradient Boosting (Setting-2), demonstrated effectiveness.
• Blood basophil count emerged as a top predictor for CRS classification.
• In Setting-2, high C-reactive protein, high blood basophil count, chronic eosinophilic pneumonia, low CT scores, and low NP scores predicted AEC ≥ 1500/µL.

Conclusions:

• ML classification highlights the role of basophils in peripheral blood eosinophilia within CRS.
• Specific features for high AEC CRS suggest considering eosinophilic CRS with unique comorbidities, even with absent nasal polyps or mild CT findings.