| The objective behind this thesis work was to continue work in developing the appropriate software for a device that is able to distinguish a healthy pediatric heart from an unhealthy one with a reasonable amount of accuracy using only the sound coming from the heart. The work for this thesis began by re-establishing the prediction performance of an artificial neural network using the same feature extraction technique that was previously used on this project by students in our lab7. This method used an FFT feature set and showed a relatively low prediction performance (65--70%), and this performance was verified.; After modifying the original model and attempting several other methods of feature extraction, wavelet analysis was implemented as a more effective feature extraction technique. With the implementation of this signal analysis tool, the performance of this new feature set improved to a higher level (75--80%). This is still below the desired accuracy, but this work shows that a more valuable feature set can be obtained using discrete wavelet decomposition instead of FFT.; Over the course of this thesis work, there have been several remarkable observations as a result of this progress: (1) Discrete wavelet decomposition is superior to the Fast Fourier Transform in dynamic signal analysis, (2) Both time and frequency features are necessary for heart sound diagnosis, and (3) Wavelet analysis is an effective tool for reducing the negative effect of noise on signal identification. |
