Noninvasive predictor of HeartMate XVE pump failure by neural network and waveform analysis

Patients increasingly require longer durations of left ventricular assist device (LVAD) therapy. Despite a recent trend toward continuous flow VADs, the HeartMate XVE is still commonly used, but its longevity remains a significant limitation. Existing surveillance methods of pump failure often give inconclusive results. XVE electrical current waveforms were collected regularly (2001-2008) and sorted into quartiles according to number of days until pump failure (Q1, 0-34; Q2, 34-160; Q3, 160-300; and Q4, 300-390 days). Thoratec waveform files were converted into text files. The 10-second electrical current, voltage waveform was identified and isolated for analysis. Waveforms were analyzed by principal component analysis (PCA) and with a fast Fourier transform. Quartiles were compared with analysis of variance (ANOVA). Waveforms (n = 454) were collected for 21 patients with failed pumps. An artificial neural network was used to predict pump failure within 30 days from the waveform characteristics identified though signal processing.