Evaluation of suction detection during different pumping states in an implantable rotary blood pump

In recent times, the problem of noninvasive suction detection for implantable rotary blood pumps has attracted substantial research interest. Here, we compare the performance of various suction indices for different types of suction and non-suction events based on pump speed irregularity. A total of 171 different indices that consist of previously proposed as well as newly introduced suction indices are tested using regularized logistic regression. These indices can be classified as amplitude based (derived from the mean, maximum, and minimum values of a cycle), duration based (derived from the duration of a cycle), gradient based (derived from the first order as well as higher order differences) and frequency based (derived from the power spectral density). The non-suction event data consists of ventricular ejection with or without arrhythmia and intermittent and continuous non-opening of the aortic valve. The suction event data consists of partial ventricular collapse that occurs intermittently as well as continuously with or without arrhythmia. In addition, we also attempted to minimize the usage of multiple indices by applying the sequential forward floating selection method to find which combination of indices gives the best performance. In general, the amplitude-based and gradient-based indices performed quite well while the duration-based and frequency-based indices performed poorly. By having only two indices ([i] the maximum gradient change in positive slope; and [ii] the standard deviation of the maximum value in a cycle), we were able to achieve a sensitivity of 98.9% and a specificity of 99.7%.