Left Ventricular Assist Device Thrombus Detection Using Wavelets and Image Classification

Cardiac related diseases are a common health risk for adults. Consequently, therapies exist to treat these aliments such as medication through heart transplant. Heart transplant, however, remains the gold standard for treating severe heart failure, but left ventricular assist devices (VADs), a cardiac blood pump, are gaining popularity and not just as bridge to surgery. These devices work in parallel with the heart, providing support to either the pulmonary or systemic circulatory circuit. VADs have been very successful in treating patients with heart failure who are unable to receive a heart transplant.

Unfortunately, with the benefits of these devices come risk of clot formation. These occlusions can cause strokes, further cardiac damage, or even death. Therefore, it is critical that these occlusions be detected at the onset. The VAD's in vivo nature makes detecting faults or occlusions difficult with traditional methods such as blood tests or sonography. In addition, these detection methods require the patient to travel to a facility equipped with such devices. Early detection of these conditions would substantially increase the VAD patient's level of care.

This work presents a method to non-invasively monitor the condition of a Thoratec HeartMate II (HMII) ventricular device. A fully functioning hemodynamic simulator was designed and fabricated at the Rochester Institute of Technology and replicates a variety of conditions in the left side of the human cardiovascular system, including pulsatile blood flow and a wide variation in patient-to-patient blood pressure conditions. A molded silicone volume within a rigid-walled sealed polycarbonate shell is used to represent the left ventricle of the heart. Pulsatile flow is generated by pneumatically controlling the air pressure between the cylinder wall and the silicone membrane. Flow is then ejected through a molded silicone, tricuspid aortic valve. Flow continues through two compliance elements which represent the resistance and compliance of the vasculature, both of which can be manually adjusted to simulate physiological conditions. Flow then returns to the atrium and through a bi-leaflet mitral valve to return to the left ventricle. The simulator has been augmented to replicate a typical VAD patient by installing a HMII - VAD within the left ventricle chamber and connecting the pump discharge to the aorta bypassing the aortic valve. The entire VAD system is encased in a soft and compliant polymer to represent human tissue and provide an accurate medium for acoustic measurement transmission. A digital stethoscope serves as the primary measurement device for this work.  

Acoustic data is collected with the digital stethoscope under a variety of simulated human cardiac conditions.  Seeded occlusions of various severity are designed and installed in the simulator to represent an aortic graft obstruction where the style and level of occlusion is consistent with those observed clinically by VAD cardiologists. The data is then ensemble averaged over a number of cardiac cycles based on a simulated ECG measurement (i.e. the pneumatic pulse trigger to the VAD chamber). These averages represent one sample of a known VAD health condition. A continuous wavelet analysis is performed and the result is a normalized and downsampled greyscale image where the pixel intensity represents a single observation feature vector. The hemodynamic simulator provides hundreds of sample measurements over a variety of conditions to be used as training and validation data for the application of a neural network and classification tree architecture. When tested on digital heart sound spectra obtained from the in vitro cardiac simulator, the classification tree showed the most favorable results, outperforming a pre-existing support vector machine method by roughly 20%.