Chronic obstructive pulmonary disease screening using time-frequency features of self-recorded respiratory sounds

Abstract

Objectives: Chronic obstructive pulmonary disease (COPD) is the third leading cause of death worldwide, with up to 70% of cases remaining undiagnosed. This paper proposes a COPD screening tool based on time-frequency representation features of self-recorded respiratory sounds.

Materials and methods: Respiratory sound samples (breath and cough sounds) were extracted from COPD and asymptomatic non-COPD volunteers using a large, scientific-purpose database. We analyzed 39 time-frequency representation features of breath and cough sounds, combined with age, sex, and smoking status, using Autoencoder neural networks and random forest (RF) algorithms. We compared the performance of different breath and cough RF models built to detect COPD: one based exclusively on sound features, one based exclusively on sociodemographic characteristics, and one based on sound features and sociodemographic characteristics.

Results: Models including breathing features outperformed models exclusively based on sociodemographic characteristics. Specifically, the model combining sociodemographic characteristics and breathing features achieved an area under the curve (AUC), accuracy, sensitivity, and specificity of 0.901, 0.836, 0.871, and 0.761, respectively, in the test set, representing a substantial increase in AUC when compared to the model based exclusively on sociodemographic characteristics (0.901 vs 0.818).

Discussion: Our results suggest that a lightweight collection of the time-frequency representation features of self-recorded beathing sounds could effectively improve the predictive performance of COPD screening or case-finding questionnaires.

Conclusion: COPD screening through self-recorded breathing sounds could be easily integrated as a low-cost first step in case-finding programs, potentially contributing to mitigate COPD underdiagnosis.

Keywords: COPD screening; artificial intelligence; chronic obstructive pulmonary disease; computer-aided diagnosis; machine learning.

Figure 1.

Structure of the Autoencoders. The input and output layers are adjusted to the number of features of the breath and cough datasets (n = 39).

Figure 2.

ROC curves of the 3 models in the breath dataset. Predictions were built in the test set. Abbreviation: ROC, receiver operating characteristic.

Figure 3.

ROC curves of the 3 models in the cough dataset. Predictions were built in the test set. Abbreviation: ROC, receiver operating characteristic.

Figure 4.

ROC curves of the 3 models of the breath dataset in a sensitivity analysis excluding all samples from participants reporting to be never smokers. Predictions were built in the test set. Abbreviation: ROC, receiver operating characteristic.