Breath as a biomarker: harnessing volatile clues in interstitial lung disease diagnosis

Despite recent advancements in interstitial lung disease (ILD) diagnostics, achieving an early and accurate diagnosis remains challenging. Patients frequently endure prolonged diagnostic journeys potentially involving invasive diagnostic testing.1 This is compounded by anxiety and reduced emotional well-being that frequently accompanies diagnostic delays and uncertainty.2,3 This is particularly challenging for patients who live farther from centers with ILD expertise, and may contribute to disparities in care and delay initiation of effective therapies.4 Even after expert assessment and multidisciplinary discussion (MDD), approximately 10% of patients remain unclassifiable, often due to missing histopathologic data when surgical risk is high.5 There is growing momentum in the field toward less invasive and more accessible diagnostic approaches. We have witnessed the emergence of cryobiopsy and genomic classifiers that distinguish usual interstitial pneumonia (UIP) from non-UIP histology patterns with high specificity.6 However, these tests still require a bronchoscopy, which carries procedural risk, and nondiagnostic results may necessitate further invasive sampling. Moreover, access to these technologies and the expertise needed to conduct and interpret this testing is not universal. Advances in molecular signatures across multiple biological domains and in quantitative computed tomographic imaging have shown promise for differentiating ILD by diagnosis and for identifying prognostic and therapeutic subgroups.7-10 Breathomics, a branch of metabolomics, is one such biologic domain, which focuses on volatile organic compounds (VOCs), byproducts of physiologic and pathologic processes as well as external exposures, that are present in exhaled breath.11 The electronic nose (eNose) is a simple-to-use point-of care-testing device that employs sensors capable of capturing the overall VOC composition.11 Prior single-center studies have demonstrated its ability to distinguish ILD from healthy controls and other chronic lung diseases and to identify individual ILD subtypes.11,12 In this issue of American Journal of Respiratory and Critical Care Medicine, Formsma and colleagues evaluated the validity of eNose as a diagnostic tool for fibrotic ILDs across multiple international centers.13 Their analysis included 587 patients with idiopathic pulmonary fibrosis (IPF), connective tissue disease–related ILD, fibrotic hypersensitivity pneumonitis (fHP), idiopathic nonspecific interstitial pneumonia, unclassifiable ILD, and interstitial pneumonia with autoimmune features, as determined by each center’s MDD. They employed a machine learning data classification approach using partial least squares discriminant analysis. They trained the model on a subset of data from selected centers with subsequent validation using data from independent centers. Using this model, the investigators compared breath profiles of each individual ILD subtype against (1) all other ILDs as a composite group and (2) against each alternative subtype individually. Model performance was assessed by calculating the area under the receiver operating characteristic curve (AUC). Their findings demonstrate that these ILD subtypes can be accurately distinguished not only when compared with all other ILDs (AUC range, 0.75-0.95 in validation sets), but also from one another (AUC range, 0.83-0.93 in validation sets). Comparing the breath profiles of patients with IPF from patients with all other ILDs resulted in an AUC of 0.89 (95% CI, 0.85-0.94) in the validation set, as an example, which was similar to prior single-center findings.11 Importantly, subgroup analyses did not reveal influence of key clinical covariates including lung function and smoking status on the discriminative ability of the technology. We applaud the authors for taking this sizable next step in advancing the diagnostic landscape in ILD. This novel technology has the potential to transform the process of ILD diagnosis in coming years given its ease of use and efficiency. Widespread availability could do much to address the significant disparities that currently exist in access to ILD diagnosis. In clinical practice, we are often able to narrow the differential to a small number of most likely ILD subtypes, making the head-to-head comparisons in this study particularly informative. The authors appropriately note that, in their use of supervised machine learning, model performance is dependent on the accuracy of MDD-derived diagnoses, which have established limitations. Machine learning models could in principle be trained using biopsy-informed MDD as reference standards; however, lung biopsy is performed in a minority of patients in current ILD practice, which limits the feasibility of relying solely on histology-anchored labels for model training.14 Regardless, this approach may still support MDD decision-making in cases of low diagnostic certainty, or in settings where ILD expert–led MDD is less accessible. The authors highlight the potential for unsupervised machine learning techniques to generate patient clusters based on breath profiles, which could reveal disease phenotypes with shared pathobiological mechanisms that transcend traditional clinicopathologic categories. In this study, participants diagnosed with unclassifiable ILD were clearly distinguishable from other ILD subtypes (AUC, 0.95 [95% CI, 0.89-1.00]). This distinct breath profile is notable for a group classically considered to be a heterogeneous catch-all. We would be interested in seeing whether unsupervised methods reveal subgroups within the fHP category given the lowest reported diagnostic accuracy for this ILD subtype (fHP vs all other ILDs: AUC, 0.75 [95% CI, 0.61-0.88]), or cluster patients by shared radiographic patterns such as UIP across ILD subtypes.15 While the current study’s use of breath analysis to classify patients into existing diagnostic categories is certainly an important step forward, the future capabilities of eNose may prove to be even more compelling. In the age of precision-based medicine, with a growing emphasis on disease behavior to guide classification and management of fibrotic ILD, highly anticipated is the ability of breathomics to assist with prediction of disease progression and response to existent and emerging pharmacotherapies. Furthermore, combining breathomics with other -omics data is a very powerful proposition remaining to be explored. The idea of using eNose technology to inform medical decision-making in real time, though unlikely to be widely available for clinical practice in the immediate future, is certainly compelling. Further demonstration of feasibility and uptake in different practice settings will be crucial to wide adoption of this method. We look forward to the proposed future work relating to whether eNose can improve ILD diagnostic confidence and provide an alternative to invasive approaches in a patient population with a decidedly elevated perioperative risk. Multimodal integration and translation into clinical practice have the potential to transform both patient care and research in this vulnerable population. Supplementary material is available at American Journal of Respiratory and Critical Care Medicine online. Please see the ICMJE disclosure forms, which have been provided as supplementary material.