The Problem
Left ventricular assist devices (LVAD) provide lifesaving therapy for patients with advanced heart failure, but there is a high rate of complications with nearly 60% of patients readmitted to hospital within one year of therapy. Strategies to predict complications could potentially reduce hospitalizations and costs related to LVAD therapy. We proposed using a novel acoustic surveillance strategy to develop a predictor of impending LVAD complications using digital stethoscope recordings of ambulatory patients on LVAD therapy.
Our Solution
We have created the largest repository of acoustic data from an ambulatory cohort of LVAD patients and paired this with clinical and event data. 24 subjects were enrolled in the study, 18 with Heartmate 3 (HM3) and 6 with Medtronic HVAD. 16 events were identified among the 24 subjects at the time of this report (16/24 = 67%), with current total followup time ranging from 4-6 months among subjects. Weekly participation rates for survey completion and acoustic recordings were consistently >80%. The study team collected LVAD acoustic recordings every 3 months at routine LVAD clinic appointments, which limits the ability to report on comprehensive results at the time of this report. For sound analysis, the signals were first downsampled and band pass filtered to restrict the frequency content of the signals to less than 500Hz, and adaptive filtering was used to isolate LVADspecific frequency components and better emphasize native heart sounds. The frequency content of the signals was analyzed pre- and post-adaptive filtering by estimating power spectral densities (PSDs) of five-second signal segments. In general, frequency peaks were observed in the first harmonic and multiples of the fourth harmonic of the pump frequency in the HVAD models, while peaks were observed at multiple harmonics of the pump frequency in the HM3 models; these findings are consistent with expectations based on the pump rotational frequency and number of blades. Baseline acoustic spectra for the 24 enrolled subjects after attenuating the LVAD-specific frequency components (i.e., post-adaptive filtering) are shown in Figure 1. Based on the PSD estimates, four clusters of heart sounds were identified, with one heart sound signature associated specifically with HM3 pumps. Additional signal analysis is ongoing to correlate spectral features with clinical outcomes.
Impact
We have had the opportunity to analyze event data from one patient at the time of this report. Please see figures 2a and 2b. The patient was admitted with ventricular arrhythmia. In the week prior to the event, there was a noticeable increase in the ratio of the peak amplitude of the second to first harmonic of the pump frequency.
This was observed with and without adaptive filtering. The lead time from observation of change in acoustic ratio was 6 days. In particular, the cause of the subject’s ventricular tachycardia was thought to be related to septal/cannula interaction and possible hypovolemia. The LVAD speed was reduced to minimize change of interaction. The patient had another defibrillator discharge to treat ventricular tachycardia several months after this initial episode, though the acoustic data from this event has yet to be collected at the 6 month followup appointment. We were able to observe changes in acoustic spectra in a patient with LVAD complication with adequate lead time for intervention to prevent complications. We similarly anticipate identifying changes from baseline acoustic spectra as we explore sound recordings from other patients who have had LVAD complications. We have created the largest repository of longitudinal LVAD acoustic data reported in the literature. We and other groups have previously reported that LVAD thrombosis is associated with changes in acoustic spectra, and our work builds on this by showing that other LVAD complications can potentially be identified and perhaps predicted using an acoustic surveillance strategy. We will leverage this robust dataset to create predictors of various LVAD complications. In addition to drafting a manuscript with our baseline data, we are currently in the process of drafting several grants for ongoing research in this field, including foundation grants, NIH R03 to develop methodology, and an NIH R01 to build a larger validation cohort.
