Machine learning methods on exhaled volatile organic compounds for distinguishing COPD patients from healthy controls

Abstract

Exhaled volatile organic compounds (VOCs) have shown promise in diagnosing chronic obstructive pulmonary disease (COPD) but studies have been limited by small sample size and potential confounders. An investigation was conducted in order to establish whether combinations of VOCs could identify COPD patients from age and BMI matched controls. Breath samples were collected from 119 stable COPD patients and 63 healthy controls. The samples were collected with a portable apparatus, and then assayed by gas chromatography and mass spectroscopy. Machine learning approaches were applied to the data and the automatically generated models were assessed using classification accuracy and receiver operating characteristic (ROC) curves. Cross-validation of the combinations correctly predicted the diagnosis in 79% of COPD patients and 64% of controls and an optimum area under the ROC curve of 0.82 was obtained. Comparison of current and ex smokers within the COPD group showed that smoking status was likely to affect the classification; with correct prediction of smoking status in 85% of COPD subjects. When current smokers were omitted from the analysis, prediction of COPD was similar at 78% but correct prediction of controls was increased to 74%. Applying different analytical methods to the largest group of subjects so far, suggests VOC analysis holds promise for diagnosing COPD but smoking status needs to be balanced.

Breath Analysis in Real Time by Mass Spectrometry in Chronic Obstructive Pulmonary Disease

“…gas chromatography MS [15,44]) or in real time by a number of different techniques. The main advantage of MS is its selectivity (especially high-resolution instruments), which enables a wide analyte coverage as well as the structural identification of the components detectable in breath.…”

Short-Term Intra-Subject Variation in Exhaled Volatile Organic Compounds (VOCs) in COPD Patients and Healthy Controls and Its Effect on Disease Classification

“…The cohort comprised of 118 patients with COPD and 63 healthy controls (Table 1). Full information on the groups and their recruitment has been published previously and in more detail [16] together with a description of COPD severity.…”

“…Smoking status has previously been demonstrated to act as a confounding factor in COPD diagnosis; with a balanced smoking status in COPD and healthy groups leading to significantly improved classification of healthy subjects [16]. This was achieved by eliminating current smokers from the analysis and treating all samples as separate objects (although data from the same subject did not appear simultaneously in the training and testing phases).…”

“…The TD system was an Ultra unit (Markes International, Llantrisant, UK) while the GC was a 6890N (Agilent Technologies) and the MS a 5973 network mass selective detector (Agilent Technologies). Methodology has been described in full previously [16]. All three repetitions from a given individual were processed in the same TD run in batches of 7 to 10 subjects i.e.…”

“…Therefore prior to analysis, this had to be compensated for so that overall trends across and between groups could be readily compared and this drift did not influence machine learning outcomes. The normalization methods used for this have been reported previously [16]. A benefit of employing such normalization is that it removes the need for placing internal standards into the samples.…”