Related Experiment Video
Updated: May 13, 2026
Absorption of Nasal and Bronchial Fluids: Precision Sampling of the Human Respiratory Mucosa and Laboratory Processing of Samples
Published on: January 21, 2018
Prediction of phenotypes by secretory biomarkers and machine learning in patients with chronic rhinosinusitis
M Becker1, A M Kist, O Wendler
1Department of Otorhinolaryngology, Head and Neck Surgery, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Bavaria, Germany. vanessa-vivien.pesold@uk-erlangen.de.
Objective:
Chronic rhinosinusitis (CRS) has traditionally been classified phenotypically according to the presence (CRSwNP) or absence (CRSsNP) of nasal polyps. However, the phenotypic dichotomy does not represent the complexity of the disease. Current research thus focuses on identifying underlying inflammatory mechanisms and distinguishing different endotypes. The objectives of this study were 1) to identify maximally predictive non-invasive biomarkers from nasal mucus, 2) to apply machine learning algorithms to use mucus-derived biomarkers to classify phenotype, and 3) to determine the feature importance of each mucus biomarker to phenotypes.
Patients And Methods:
This is an IRB-approved study of 103 CRS patients (37 CRSsNP, 66 CRSwNP). Nasal mucus was collected using merocele sponges after a 3-week steroid washout period. The nasal mucus was then examined for twelve cytokines/inflammatory protein biomarkers, including interferon (IFN)-γ, interleukin (IL)-4, -5, -17A, -22, immunoglobulin (Ig) E, cystatin-SA (CST-2), eosinophilic cationic protein (ECP), matrix metalloproteinase-9 (MMP-9), pappalysin-A (PAPP-A), periostin, and serpin E1. Protein concentrations were determined by ELISAs and Luminex assays. For phenotype classification, different artificial intelligence algorithms in increasing complexity, including t-distributed stochastic neighbor embedding (t-SNE), Adaboost, and XGBoost, were applied to the data from the biomarker analysis.
Results:
TThe analysis showed that IL-5 is a non-invasive marker to distinguish between the two phenotypic clusters. This was true for immune cell-derived proteins, and all proteins were analyzed conjointly. Periostin and CST-2 showed the highest feature importance for the epithelial- and tissue-derived proteins. The combination of IL-5, IgE, IL-17, and periostin showed the highest accuracy for prediction.
Conclusions:
Nasal mucus can predict phenotypes similar to tissue, with IL-5 as the main trigger for clustering. Periostin and CST-2 may be part of important targetable pathways. Future efforts will be directed at determining how these markers may be used to guide therapeutic choices and individualize treatment.
