Correlating spirometry findings with auscultation sounds for diagnosis of respiratory diseases

Abstract

Spirometry is the pulmonary function test (PFT) used for the diagnosis and severity measurement of respiratory illness. Although it is a non-invasive procedure still requires expertise, patient cooperation and repeated maneuver. Auscultation is the conventional method doctors use in the clinical environment. It can be used for early, efficient and remote diagnosis of respiratory diseases. In this paper, a method is proposed which can classify auscultation sounds collected from a simple stethoscope in OPD settings into three major categories, i.e., healthy, obstructive and restrictive which is traditionally diagnosed using spirometry. The technique proposed in this work is based on pre-processing respiratory sounds by enhancing them using specifically designed parametric equalizer filter. Spectro-temporal Gabor filter-bank-based features (GBFB) are extracted from pre-processed respiratory sounds to classify them using different machine learning models. Also, an exhaustive study is performed to determine most efficient features for classifying respiratory sounds. The proposed method is novel and gives state-of-the-art results with an F1 score of 97.95%, precision of 100% and recall of 96%. The method proposed can be used for automated diagnosis of respiratory illness and can circumvent spirometry test in some scenarios. It can even help people living in remote and rural areas where finding an expert to carry out the PFT is limited.