Clinical Utility of Machine Learning Methods Using Regression Models for Diagnosing Eosinophilic Chronic Rhinosinusitis.

Objective: Machine learning methods using regression models can predict actual values of histological eosinophil count from blood eosinophil levels. Therefore, these methods might be useful for diagnosing eosinophilic chronic rhinosinusitis, but their utility still remains unclear. We compared 2 statistical approaches, and investigated the utility of machine learning methods for diagnosing eosinophilic chronic rhinosinusitis.

Study design: Retrospective study.

Setting: Medical center.

Methods: Data, including eosinophilic levels, obtained from blood and sinonasal samples of 264 patients with chronic rhinosinusitis (257 with and 57 without nasal polyps) were analyzed. We determined factors affecting histopathological eosinophil count in regression models. We also investigated optimal cutoff values for blood eosinophil percentages/absolute eosinophil counts (AECs) through receiver operating characteristic curves and machine-learning methods based on regression models. A histopathological eosinophil count ≥10/high-power field was defined as eosinophilic chronic rhinosinusitis.

Results: Blood eosinophil levels, nasal polyp presence, and comorbid asthma were factors affecting histopathological eosinophil count. Cutoffs between the 2 statistical approaches differed in the group with nasal polyps, but not in one without nasal polyps. Machine-learning methods identified blood eosinophil percentages ≥1% or AEC ≥100/μL as cut-offs for eosinophilic chronic rhinosinusitis with nasal polyps, while ≥6% or ≥400/μL for one without nasal polyps.

Conclusion: Cut-offs of blood eosinophil levels obtained by machine-learning methods might be useful when suspecting eosinophilic chronic rhinosinusitis prior to biopsy because of their ability to adjust covariates, dealing with overfitting, and predicting actual values of histological eosinophil count.