|About GSBS | FAQ | Job Opportunities | Search UMDNJ|
M.S., Biomedical Engineering
Drexel University - 2007
Thesis Advisor: John L. Semmlow, Ph.D.
Graduate Program in Biomedical Engineering
R.U. Biomedical Engineering Building
Room 126, Busch Campus
Friday, September 23, 2011
Coronary artery disease (CAD) is a long term process in which the arteries that deliver blood and oxygen to the tissues of the heart become occluded. This is the number one cause of death in the United States. Unfortunately, CAD is typically asymptomatic until there is significant disease progression and current diagnostic techniques such as the coronary angiogram or CT angiogram are expensive and invasive. For these reasons CAD may go undiagnosed until it is too late.
The acoustic approach to diagnosis of CAD is inexpensive, safe, and easy to administer, which would allow for its use as an early screening tool. However, the acoustic method has never shown the ability to produce clinically relevant accuracy. In addition, the properties of the acoustic signal have not been fully characterized, as the signal is of very low power and buried in considerable noise. In this study, I develop an acoustic-based system for diagnosing CAD using nonlinear signal processing. This study contains three main thrusts: evaluation of hardware and data acquisition protocol; development of nonlinear signal processing methods; and data acquisition, classification, and assessment of CAD diagnostic capability on patients.
Hardware provided by SonoMedica, Inc. was evaluated for noise and spectral characteristics and improved through a modest amplifier redesign. The new design was seen to lower the baseline electronic noise, improve the signal-to-noise ratio (SNR), and reduce signal clipping (analog signal saturation). A data acquisition protocol was designed which maximized SNR. Signal SNR was found to be most sensitive to the level of tension applied to a strap used to hold microphones on the chest. Heart sound recordings were made using this protocol on 16 normal and 15 diseased subjects. Diastolic signal segments were isolated from these recordings for subsequent signal processing.
Signal processing methods included chaos-based dynamical analysis, entropy-based methods, and instantaneous frequency analysis. A new entropy method, termed path length entropy, PLE, was developed and evaluated. In addition a well-known spectrally-based algorithm for determining entropy was modified to account for shortcomings in this method. Other algorithms evaluated included approximate entropy (ApEn), sample entropy (SampEn), multiscale entropy (MSE), permutation entropy and mutual information function. Instantaneous frequency was evaluated using the Hilbert Huang transform, the normalized Hilbert-Huang transform, and numerical quadrature. Nonlinear dynamical analysis included an estimation of maximum Lyapunov exponent and correlation dimension.
Path length entropy is a predictability measurement based on the likelihood of a change of direction in a signal trajectory. PLE was motivated by the observation of 1/f structure in our measured signals. Path length entropy was the most effective measurement with a sensitivity-specificity of 80%-81%. The next highest accuracy measurements were Hurst exponent and modified spectral entropy. Both provided a sensitivity-specificity of 80%-75%. Entropy- based measurements were the most effective, followed by instantaneous frequency- based methods and lastly, chaos- based methods. PLE and modified spectral entropy analysis suggested that normal subjects, rather than diseased, have signals which contain less regularity.
Surrogate data analysis indicated the presence of nonlinearity within diastolic segments, but it was weak. Digital filtering was shown to improve the accuracy of measurements, likely due to improved SNR. Finally, a short- term PLE analysis did not provide evidence for the existence of discrete nonlinear CAD events that would give rise to bruits.
The research presented here has refined a system for measuring diastolic sounds from normal and diseased subjects, used that system to acquire data from both diseased and normal subjects, developed a new nonlinear signal processing algorithm, and evaluated the ability of the new algorithm along with existing nonlinear algorithms to differentiate between normal and diseased subjects. This research has also presented evidence for the presence of nonlinear behavior in diastolic segments and the utility of using nonlinear algorithms to detect CAD from heart sound recordings