Exploring the Ability of Electronic Nose Technology to Recognize Interstitial Lung Diseases (ILD) by Non-Invasive Breath Screening of Exhaled Volatile Compounds (VOC): A Pilot Study from the European IPF Registry (eurIPFreg) and Biobank.

Abstract

Background: There is an increasing interest in employing electronic nose technology in the diagnosis and monitoring of lung diseases. Interstitial lung diseases (ILD) are challenging in regard to setting an accurate diagnosis in a timely manner. Thus, there is a high unmet need in non-invasive diagnostic tests. This single-center explorative study aimed to evaluate the usefulness of electronic nose (Aeonose®) in the diagnosis of ILDs. Methods: Exhaled volatile organic compound (VOC) signatures were obtained by Aeonose® in 174 ILD patients, 23 patients with chronic obstructive pulmonary disease (COPD), and 33 healthy controls (HC). Results: By dichotomous comparison of VOC’s between ILD, COPD, and HC, a discriminating algorithm was established. In addition, direct analyses between the ILD subgroups, e.g., cryptogenic organizing pneumonia (COP, n = 28), idiopathic pulmonary fibrosis (IPF, n = 51), and connective tissue disease-associated ILD (CTD-ILD, n = 25) were performed. Area under the Curve (AUC) and Matthews’s correlation coefficient (MCC) were used to interpret the data. In direct comparison of the different ILD subgroups to HC, the algorithms developed on the basis of the Aeonose® signatures allowed safe separation between IPF vs. HC (AUC of 0.95, MCC of 0.73), COP vs. HC (AUC 0.89, MCC 0.67), and CTD-ILD vs. HC (AUC 0.90, MCC 0.69). Additionally, to a case-control study design, the breath patterns of ILD subgroups were compared to each other. Following this approach, the sensitivity and specificity showed a relevant drop, which results in a poorer performance of the algorithm to separate the different ILD subgroups (IPF vs. COP with MCC 0.49, IPF vs. CTD-ILD with MCC 0.55, and COP vs. CT-ILD with MCC 0.40). Conclusions: The Aeonose® showed some potential in separating ILD subgroups from HC. Unfortunately, when applying the algorithm to distinguish ILD subgroups from each other, the device showed low specificity. We suggest that artificial intelligence or principle compound analysis-based studies of a much broader data set of patients with ILDs may be much better suited to train these devices.