Uncovering Key Features for Predicting Comorbid Chronic Eosinophilic Pneumonia in Chronic Rhinosinusitis via Machine Learning

This is the authors' abstract. We don't add key points for this paper.

Background Chronic eosinophilic pneumonia (CEP) can occur concurrently with chronic rhinosinusitis (CRS). However, crucial features of comorbid CEP in patients with CRS remain unclear. Methods Features of comorbid CEP were thoroughly investigated using machine learning (ML). In ML, (i) highly predictable performance and (ii) high interpretability (e.g., presenting classification rules understandable to clinicians) are two objectives with a tradeoff relationship, resulting in both being simultaneously unachievable by a single ML model. In this study, for (i), ML models were examined to check their predictive performance, and for (ii), decision tree (DT) was used. In addition, to address the lack of interpretability in (i), SHapley Additive exPlanations (SHAP) was applied. Results In total, 372 CRS samples (21 with CEP) were collected. In the CRS with CEP group, 19 patients were diagnosed with eosinophilic CRS (ECRS). In (i), extreme gradient boosting (XGBoost)/random forest (RF) showed a higher AUC (area under the ROC (receiver operating characteristic) curve) than logistic regression/support vector machine. In (ii), the top feature was a blood eosinophil count ≥ 1446/µL, followed by a white blood cell (WBC) ≥ 9.25 × 10 3 /µL, and C‐reactive protein (CRP) ≥ 0.335 mg/dL. SHAP, based on XGBoost and RF, selected elevations in the blood eosinophil count, CRP, and WBC count as the top three features. Conclusion DT and SHAP selected the same three top features of CRS with CEP. When patients with CRS satisfy the DT algorithm, they may have ECRS with CEP. Therefore, otolaryngologists should perform sinonasal biopsies and chest imaging.