Classification of heart valve sounds from experiments in an anechoic water tank

In vivo studies in both sheep and humans were plagued by a number of problems including movement artifacts, biological noise, low signal-to-noise ratio (SNR), chest-wall reverberation, and limited bandwidth recordings as discussed by [1]. To overcome these problems it was decided to record heart valve sounds under controlled conditions deep in an anechoic water tank, free from reverberation noise. The main goal of this experiment was to obtain measurements of ''pure'' heart valve sounds free of the scattering effects of the body. Experiments were conducted at the Transdec facility in San Diego [2]. We used a high quality hydrophone together with a wide-band data acquisition system [2]. We recorded sounds from 100 repetitions of the opening-closing cycles on each of 50 different heart valves, including 21 SLS valves and 29 intact valves. The power spectrum of the opening and closing phases of each cycle were calculated and outlier spectra removed as described by Candy [2]. In this report, we discuss the results of our classification of the heart valve sound measurements. The goal of this classification task was to apply the fundamental classification algorithms developed for the clinical data in 1994 and 1996 to the measurements from the anechoic water tank. From the beginning of this project, LLNL's responsibility has been to process and classify the heart valve sounds. For this experiment, however, we processed both the opening sounds and closing sounds for comparison purposes. The results of this experiment show that the classifier did not perform well because of low signal-to-noise ratio and excessive variability in signal power from beat-to-beat for a given valve