A Research Of Neural Network Modeling And Disease Prediction Of Heart Sound Signal

As the increasing demands for health of people,getting warnings and seeking for medical attention in time when the body may be abnormal is a good way to stay healthy.However,most people do not assess their own health status and can not understand the symptoms of changes of their physical health status in time.The best way of treating diseases is timely prevention when symptoms appear.Therefore,an important research is letting the machine learn the features of health indicators and timely alert when these features is getting bad.The heart sound is an important part of health indicators,which is also easy to collect.Heart sound represents the health of heart.Determining whether a heart sound indicates a healthy heart through artificial neural networks is a significant part of health monitoring.This thesis studies the characteristics of heart sound signals and constructed a neural network model for prediction The main work of this paper are as follows:1.Propose a method of heart sound noise reduction based on Butterworth filter and wavelet transform.Heart sound signals concentrate in the low frequency region and have centralized energy.Low pass Butterworth filter gives low frequency part of heart sound,and wavelet transform focus on energy.2.Propose a method of feature extraction for heart sound signal based on Mel frequency cepstrum coefficient.Re-sampling heart sound according to its periodicity.Then calculating its MFCC in order to ensure that a period of heart sound signal is framing into the same count of frames.Combining its period with MFCCs of two periods of signal which starts from S1 heart sound as feature vector.3.Constructed a heart health prediction model based on a deep neural network model.The features extracted from the heart sound signals are used to predict heart health and evaluate the performance of the neural network.Then,optimizing the network model to make it more suitable for the extracted features.The comparison with the prediction results of common classifiers and statistical methods shows that the performance of the neural network constructed in this thesis is more stable than those methods,and the performance has also been partially improved.