This research project attempts to compare traditional statistical techniques, expert systems, and artificial neural networks in diagnosing pediatric heart sounds. An additional component of this project is to capture the heart sounds with a portable computer and make them available over the Internet. Co-investigator for this research is Dr. E. Bayne (Department of Pediatrics, University of Florida Medical School at Jacksonville).

Pilot work for the study has been completed (Erik Hudson's thesis). A microphone with a special pre-amp was attached to a stethoscope. The microphone was used as the input device to a ProAudio sound card and the resulting WAV file saved for later analysis. The WAV file was then used as data for linear regression (nonlinear regression will be used later) and a backpropagation trained neural network. The regression model was not able to identify any of the heart sounds. The artificial neural network was able to accurately determine healthy and abnormal heart sounds 74.2% of the time from a data set not used in training.

Work is now underway to extend the pilot study. Nonlinear statistical techniques and additional neural network architectures are being studied to improve detection accuracy.