Noninvasive diagnosis of mechanical failure of the implanted total artificial heart using neural network analysis of acoustic signals.

Abstract

Acoustic signal measurement has been proposed as a noninvasive method of detecting mechanical failure of the implanted total artificial heart. However, differences in acoustic spectra obtained from undamaged and damaged devices may be difficult to distinguish using standard techniques, such as visual inspection or statistical analysis. A new technique, artificial neural network analysis, which has been used successfully on other problems of pattern recognition and classification, was applied to improve the detectability of the acoustic method. Acoustic signals were measured using two different devices in one damaged and one undamaged electrohydraulic total artificial heart, both in a mock circulation set-up and in animal experiments where they were implanted in eight post mortem sheep and the acoustic signal measured using a microphone placed at the skin surface. Spectra of the acoustic waveforms were calculated by discrete Fourier transformation and 400 values (representing the log magnitude in each 2.5 Hz band of the spectrum between 0 and 1 kHz) and used as input to the neural network. A three layer backpropagation neural network containing 400 input nodes, 20 intermediate nodes, and one output node was able to forms. The trained neural network then perfectly distinguished damaged waveforms from undamaged ones, with good separability. Because the neural network's output can take on a value between two extremes denoting damaged and undamaged states, it is possible to detect any progressive failure at relatively earlier stages. With multiple output node configuration, it could also classify the different types of damage using single acoustic signal waveforms.