Machine learning of biomarkers and clinical observation to predict eosinophilic chronic rhinosinusitis: a pilot study

Background: Subtyping chronic rhinosinusitis (CRS) by tissue eosinophilia has prognostic and therapeutic implications, and is difficult to predict using peripheral eosinophil counts or polyp status alone. The objective of this study was to test machine learning for prediction of eosinophilic CRS (eCRS). Methods: Input variables were defined as peripheral eosinophil count, urinary leukotriene E4 (uLTE4) level, and polyp status. The output was diagnosis of eCRS, defined as tissue eosinophil count >10 per high-power field. Patients undergoing surgery for CRS were retrospectively reviewed for complete datasets. Univariate analysis was performed for each input as a predictor of eCRS. Logistic regression and artificial neural network (ANN) machine learning models were developed using random and surgeon-specific training/test datasets. Results: A total of 80 patients met inclusion criteria. In univariate analysis, area under the receiver operator characteristic curve (AUC) for peripheral eosinophil count and uLTE4 were 0.738 (95% confidence interval [CI], 0.616 to 0.840) and 0.728 (95% CI, 0.605 to 0.822), respectively. Presence of polyps was 94.1% sensitive, but 51.7% specific. Logistic regression models using random and surgeon specific datasets resulted in AUC of 0.882 (95% CI, 0.665 to 0.970) and 0.945 (95% CI, 0.755 to 0.995), respectively. ANN models resulted in AUC of 0.918 (95% CI, 0.756 to 0.975) and 0.956 (95% CI, 0.828 to 0.999) using random and surgeon-specific datasets, respectively. Model comparison of logistic regression and ANN was not statistically different. All machine learning models had AUC greater than univariate analyses (all p < 0.003). Conclusion: Machine learning of 3 clinical inputs has the potential to predict eCRS with high sensitivity and specificity in this patient population. Prospective investigation using larger and more diverse populations is warranted.