APPLYING MACHINE LEARNING ALGORITHMS WITH VOICE PATTERN ANALYSIS TO THE DIAGNOSIS OF CORONARY ARTERY DISEASE
DIGITAL MEDICINE & WEARABLE TECHNOLOGY
Author: Sharon Kim
Status: Project Concept
The American Heart Association’s 2017 Heart Disease and Stroke Statistics shows that cardiovascular disease (CVD) accounts for nearly 801,000, or 1 of every 3, deaths in the US. The leading type of cardiovascular disease is Coronary Artery Disease (CAD) which contributes to 45.1% of cardiovascular disease related deaths. For people who have silent CHD-where no signs and symptoms occur-CAD is impossible to diagnose before heart attack, heart failure, or arrhythmia occurs. A study by Abdel-Ellah et al observed that age, gender, diabetes, and smoking were strong significant predictors for CAD. The predictors for CAD are widely known throughout the world of medicine, but screening for CAD does not occur until symptoms like angina appear. This is because the patient must engage in a diagnostic test like the ECG or holter monitor which may be invasive or unwieldy for patients and health care providers. However, voice signal analysis is a new type of diagnostic measure that may collect patient data indefinitely and be minimally invasive to patient lifestyle. A study by Elad et al observed the correlation between speech patterns and CAD and determined that a specific cycle time of fluctuations in frequency during monitored speech correlated with a 2.4 increased likelihood of CAD. The univariate binary logistic regression analysis identified 11 voice features that were associated with CAD. Based on the results of this research, building a machine learning algorithm that analyzes voice signals to diagnose for CAD seems to be an immediate possiblity. The algorithm can be programmed into a speech monitor and used as a wearable device to screen patients. Using a Multi Dimensional Voice Program (MDVP) that can calculate as many as 33 acoustic parameters from a voice sample can be applied to the algorithm to yield more specific and sensitive results in predicting CAD. Because audio data is in an unstructured format, one limitation to voice pattern analysis includes an under-exploited opportunity in the field of unstructured data which means more research must be done in this field to validate our algorithm; another limitation may be that patients could be wary of having their privacy invaded through speech monitoring. Though this may still be a relatively new diagnostic tool compared to the ECG, the implications for creating algorithms that analyze voice patterns opens the possibility for diagnosing not only CAD but other diseases. Further research into other body systems that may have a relationship with speech and the causation of disease and altered speech are needed in order for health care providers and legislators to have confidence to utilize speech as a diagnostic tool clinically.