RESUMO
BACKGROUND: Reliable noninvasive methods are needed to identify endotypes of chronic rhinosinusitis with nasal polyps (CRSwNP) to facilitate personalized therapy. Previous computed tomography (CT) scoring system has limited and inconsistent performance in identifying eosinophilic CRSwNP. We aimed to develop and validate a radiomics-based model to identify eosinophilic CRSwNP. METHODS: Surgical patients with CRSwNP were recruited from Tongji Hospital and randomly divided into training (n = 232) and internal validation cohort (n = 61). Patients from two additional hospitals served as external validation cohort-1 (n = 84) and cohort-2 (n = 54), respectively. Data were collected from October 2013 to May 2021. Eosinophilic CRSwNP was determined by histological criterion. The least absolute shrinkage and selection operator and the logistic regression (LR) algorithm were used to develop a radiomics model. Univariate and multivariate LR were employed to build models based on CT scores, clinical characteristics, and the combination of radiological and clinical characteristics. Model performance was evaluated by assessing discrimination, calibration, and clinical utility. RESULTS: The radiomics model based on 10 radiomic features achieved an area under the curve (AUC) of 0.815 in the training cohort, significantly better than the CT score model based on ethmoid-to-maxillary sinus score ratio with an AUC of 0.655. The combination of radiomic features and blood eosinophil count had a further improved performance, achieving an AUC of 0.903. The performance of these models was confirmed in all validation cohorts with satisfying predictive calibration and clinical application value. CONCLUSIONS: A CT radiomics-based model is promising to identify eosinophilic CRSwNP. This radiomics-based method may provide novel insights in solving other clinical concerns, such as guiding personalized treatment and predicting prognosis in patients with CRSwNP.