A deep learning approach for chronic obstructive pulmonary disease diagnosis from human exhaled breath gases

A proficient real-time decision support system has the potential to reduce the daily probability of acute exacerbation and loss of control for those suffering from chronic obstructive pulmonary disease (COPD). Applying statistical learning techniques to well-structured, medical E-nose data typically results in high accuracy. Volatile organic compounds or changes by disease processes can be measured in exhaled breath.This work elaborated on the integration of sensors into a sensor array, sampling methodologies, and an algorithm for data analysis. The clinical feasibility of the device was assessed in 40 COPD patients, 20 controls, 8 smokers, and 10 ambient air samples. The classification model utilizing Bi-Directional Long Short-Term Memory (Bi-LSTM) achieved an accuracy, sensitivity, specificity, and area under the curve of 99%, with recall, precision, and F1-score of 1 for COPD classification. The gas sensor array was non-invasive, economical, and provided a quick response. Research has shown that the VOC profiles of COPD patients differ from those of healthy controls, indicating that the E-nose system may serve as a viable diagnostic tool for COPD patients.