Background: Previous studies have been limited to the utility of clinical features and invasive nasal mucosal biomarkers in the prediction of chronic rhinosinusitis (CRS) outcomes. This study aimed to identify noninvasive biomarkers associated with difficult-to-treat CRS, enabling physicians to subgroup patients into risk groups for poor outcome before surgery. Methods: Three hundred and nine CRS patients undergoing endoscopic sinus surgery were finally enrolled. Patients treated with oral or intranasal glucocorticoids within 3 months or 1 month before surgery, respectively, were excluded. Baseline clinical characteristics, nasal secretions and peripheral blood samples were collected before surgery. The protein levels of 39 biological markers were detected by the Bio-Plex suspension chip method. Classification and regression tree analysis was applied to establish prediction model for difficult-to-treat CRS determined one year after surgery. A random forest algorithm was used to confirm the discriminating factors that formed the classification tree. Results: In the cohort with nasal secretion sample (n = 189), 21% of CRS patients were diagnosed as difficult-to-treat after 1 year of follow-up. Nasal secretion CCL17 level, hyposmia score, allergic rhinitis comorbidity, and nasal secretion MIP-1 beta level were found important predictors of difficult-to-treat CRS. A classification tree separated patients into 5 subgroups leading to an overall predictive accuracy of 94%. However, none of the plasma biological markers were associated with difficult-to-treat CRS in the cohort with blood sample (n = 128). Conclusions: Patients with difficult-to-treat-CRS were characterized by higher nasal secretion levels of CCL17 and MIP-1 beta, severe hyposmia and concomitant allergic rhinitis. The classification tree could be useful to identify patients with high risk of poor outcome prior to surgery and offer more personalized interventions. However, since only patients without preoperative steroid treatments were included in this study, the generalization of our predictive model in other patient populations should be considered with caution.