Detection of Coronary Artery Disease Using Automutual Information

Coronary artery disease (CAD) is the leading cause of death in the United States, but there is no detection method that is suitable as an early screening tool. The exercise stress test can be performed noninvasively but requires physical exertion and has low accuracy. A nuclear stress test has higher accuracy, but requires the use of a radio-nucleotide that exposes the patient to radiation. Detection of CAD using acoustic signals recorded from the chest is inexpensive, uses no radiation and is noninvasive. Therefore, the acoustic method is an ideal early screening tool. CAD sounds (bruits) are produced by turbulent blood flow caused by partially obstructed arteries and turbulence is a nonlinear process. Therefore, this study evaluated a nonlinear signal analysis method, the automutual information function (AMIF(τ)), for detection of CAD bruits. The AMIF(τ) was applied to diastolic signals from 16 normal and 15 diseased subjects. Four parameters of the AMIF(τ) curve were evaluated: (1) lag ∈ the lag when the value of the curve is 1/e; (2) the lag of the first AMIF(τ) peak; (3) the value of the first AMIF(τ) peak; and (4) the integral of the curve. Using the standard deviation (std) and mean value of AMIF(τ) parameters as features in a linear support vector machine classifier, the best individual AMIF(τ) parameter was lag ∈ This classified normal and diseased subjects retrospectively with a sensitivity-specificity of 80-69%, with 74% accuracy. Using a combination of AMIF(τ) measurements further improved accuracy. The combination of std values for first peak lag and lag ∈ gave a sensitivity-specificity of 87-75% with an 81% accuracy. Since the mutual information function is similar to the linear autocorrelation, the autocorrelation was also considered, using the same parameters to quantify the correlation function. The best combination of autocorrelation parameters, the first peak lag and mean of the sum of the autocorrelation function, gave an accuracy of 75%. The best individual autocorrelation measurement was the first peak lag, with a sensitivity-specificity of 67-75% and an accuracy of 71%. These results indicate AMIF(τ) is sensitive to signatures of CAD bruits and may be useful for acoustic detection of CAD. However, linear correlation analysis discriminated normal and diseased subjects nearly as well, indicating that the acoustic signals recorded contain mostly linear information. © 2012 Biomedical Engineering Society.