An integral component in the development of a control strategy for implantable rotary blood pumps is the task of reliably detecting the occurrence of left ventricular collapse due to overpumping of the native heart. Using the noninvasive pump feedback signal of impeller speed, an approach to distinguish between overpumping (or ventricular collapse) and the normal pumping state has been developed. Noninvasive pump signals from 10 human pump recipients were collected, and the pumping state was categorized as either normal or suction, based on expert opinion aided by transesophageal echocardiographic images. A number of indices derived from the pump speed waveform were incorporated into a classification and regression tree model, which acted as the pumping state classifier. When validating the model on 12,990 segments of unseen data, this methodology yielded a peak sensitivity/specificity for detecting suction of 99.11%/98.76%. After performing a 10-fold cross-validation on all of the available data, a minimum estimated error of 0.53% was achieved. The results presented suggest that techniques for pumping state detection, previously investigated in preliminary in vivo studies, are applicable and sufficient for use in the clinical environment.