Voice Patterns for Classification and Regression of Atherosclerotic Coronary Disease

Abstract

Abstract Background: Cardiovascular diseases are the main cause of mortality in both genders, being coronary artery disease the most prevalent type. Risk factors provide a limited help to estimate the presence of disease, acknowledging the need to investigate new techniques or biomarkers. In this study, it is our aim to evaluate the feasibility of using acoustic patterns of European Portuguese to infer about coronary disease; Methods: By collecting audio signals from patients diagnosed with heart disease and healthy subjects, a new database was developed, consisting of audio recordings and clinical metadata from a total of 84 participants. Using a combination of acoustic features, risk factors and clinical information, with distinct machine learning models, we explored binary classification and regression; Results: The Random Forests algorithm allowed to achieve a top accuracy of 88% for a binary classification (healthy vs disease) and ElasticNet allowed to achieve the minimum mean absolute error, 0.14, in a regression task. Fewer observations in higher CAD-RADS categories were limiting challenges; Conclusions: A combination of speech biomarkers and medical information can be used for identifying patterns of atherosclerotic coronary disease.