Automated Detection and Classification of Lung Sound Using 3-D Printed Stethoscope

Abstract

Auscultation is of prime significance in a clinical setting as it has the potential for accurate diagnosis of pulmonary diseases. As such effective diagnosis from heart and lung sounds is often subject to clinician's experience and his/her hearing capability. Hence adaptive noise cancellation and separation of heart and lung sounds is an area of intensive research. Conventionally doctors detect respiratory diseases through auscultation, by advising the patients to take deep breaths of their own for the distinct hearing of the chest sounds and for masking any other noise. But it is unfeasible for patients who are in a coma and for infants. In the case of doctors with less experience, it is a common practice to send the patients for different pulmonary function tests to ensure the disease. But it might take a day or more to get the results and may charge an additional expense to the patients. The paper discusses the development of a 3-D printed electronic stethoscope that can separate heart and lung sounds and support in diagnosing lung diseases in a glimpse of time. The separation of the acquired lung sounds from the stethoscope model is done using frequency-based filtering and Adaptive Line Enhancer (ALE) algorithms implemented in a Teensy 4.0 supported with an audio sheild. The obtained lung sounds are later transferred to the computer where the signals are displayed and Machine learning algorithms are employed to detect abnormal lung sounds that are indicative of different lung disorders. The conditions investigated are Wheezes, Crackle, Both, Without, Normal. The features extraction techniques which used are FFT, MFCC and Spectral imaging. The signal processing is done using Python, on the Google Colab platform. A comparative analysis of accuracy of different machine learning algorithms are carried out, namely, SVM (Support Vector Machine), KNN (K Nearest neighbor), Naive Bayes and CNN. The best model obtained was CNN with MFCC features, resulting in an overall accuracy of 70% for ICBH Database and 61.5% for hospital data samples. Since lung disorders are becoming one of the biggest killers in the world, this stethoscope will be unquestionably a handy and intelligent tool for pulmonologists.