Respiratory Sound Classification Using AI

Overview

This project is inspired from a personal experience of a delayed diagnosis and treatment of pneumonia, which partially resulted from an inconsistency in auscultation interpretation. This project used machine learning techniques to detect the presence of crackles and/or wheezes in digitally recorded respiratory audio samples. A dataset from ICBHI (International Conference on Biomedical Health Informatics), 2017 was used to develop and test machine learning models for respiratory sound and disease classification. It consists of 5.5 hours of recordings containing 6898 respiratory cycles in 920 annotated audio samples from 126 patients. The audio samples were converted to spectrograms through Fourier transformation, allowing image classification approaches to be applied. Multiple Random Forest, CNN and ResNet50 based models were developed, using techniques such as resampling, threshold optimization and transfer learning. A 4-class (healthy, crackle, wheeze, both crackle and wheeze) classification accuracy of 58.38% is achieved with a ResNet50 based model. A stethoscope attachment cell phone adapter prototype was developed. Respiratory sounds were collected from 4 participates using this prototype device, and the ML networks correctly classified the audio. ML networks were also developed for respiratory diseases diagnosis, achieving prediction accuracy of 79.71%. Using the machine learning method for respiratory sound classification and disease diagnosis offers a way for patients to conveniently self-examine, providing consistency in the interpretation of auscultation and allowing more breathing cycles to be factored into the diagnosis.

Respiratory Sound Classification Using AI

Overview

Source coverage

Awards (2)

Competition history

Resources

Related projects

Using Deep Learning to Categorize Abnormal Respiratory Sounds

A Low-Cost Computer-Aided Lung Auscultation Apparatus and Automated Diagnosis of Respiratory Illnesses

Diagnosis of Respiratory Diseases Through Physiological Sound Analysis

Diagnosing the Stage of COVID-19 using Machine Learning on Breath Sounds

Using Machine Learning to Diagnose Respiratory Illness

Device for Analysing Coughing Patterns to Diagnose and Monitor Asthmatic Patients

RESPiRE: Respiratory Evaluation via Sensory Platform in Real Time Using Edge Learning

RespiraScan: Using Breathing Patterns as a Biomarker to Diagnose Restrictive Lung Disorders

Browse more like this