Noninvasive detection of physiologically significant pumping states in an implantable rotary blood pump / Ooi Hui Lee

Heart failure is a serious health problem that could potentially be life threatening as the inflicted heart lacks the ability to supply sufficient oxygen rich blood to the rest of the body. This spurred the emergence of implantable rotary blood pump (IRBP) that is designed to provide an alternative route for blood flow as opposed to the native route that may be obstructed or problematic due to different circumstances. In particular, much interest has been garnered on the subject of pump state detection due to the potential deleterious outcomes that is associated with over-pumping. The full unloading of the left ventricle (LV) over long period of time in a pump state known as aortic valve nonopening (ANO) may cause aortic valve fusion and thrombosis. Excessive pumping in a pump state known as ventricular suction may induce several complications such as arrhythmia induction, shift of septum, tricuspidal anastomosis and dislodging of thrombi. In this study, over-pumping states such as ANO and ventricular suction are investigated by employing the pump speed signal that is acquired noninvasively from four greyhounds that consists of different levels of systemic vascular resistance (SVR) and total blood volume. A nested classification strategy is applied in two stages, with the first one involves the detection of ventricular suction whereas the second stage was focused on distinguishing ANO state from the normal ventricular ejection (VE) state. The classification task is implemented by evaluating newly introduced indices (Ran2, Ran3, Sta1, Rms1, Rms3, Rmr1, Rmr2, Rmr3) in addition to the existing indices for the different pump states. Four types of classification algorithms, namely linear discriminant analysis (LDA), logistic regression (LR) , back propagation neural network (BPNN) and k-nearest neighbor (KNN) are applied to the computed indices to assess their performance in identification of the different pump states. From the study it is observed that ventricular suction detection achieved accuracy of 94% when implemented individually using the duration index. The performance for combination of indices was noted to have improved up to 99.5% (five indices). As for ANO pump states, combination of root mean square and standard deviation has successfully performed the detection with accuracy of 93%. Further addition of indices of (five indices) will produce accuracy of 94.6%