Variable generalization performance of deep learning algorithms to classify Heartbeat Sound signal

Project Title

Project Area of Specialization

Artificial Intelligence

Project Summary

At now, heart disease is the leading cause to deaths. Concisely to takeover this situation, heartbeat sound classification is conductive way to determine heart diseases. Patterns and features extraction are the major problem to classify heartbeat sound. In this study, three heart beat sound categories are observed such as Normal, Murmur and Extra-systole. Furthermore, band filter was used to extract the noise from the heart beat signal. Therefore, we applied the fixed size of sampling rate of each sound signals. For reducing the dimensions of the frame-rate we apply down-sampling techniques to get more difference of features and minimize the dimension of the frame rate. Subsequently, we applied Random Forest (RF), Multi-layer perceptron (MLP) and Recurrent Neural Network (RNN) which are essentially based on the Long short term memory (LSTM), Dropout, Dense, and Softmax layer. On that conclusion, the proposed method is more competitive compared to other method.

Project Objectives

The objective of this study is to evaluate and achieve the best accuracy with the lowest error rate in diagnosing disease. For doing this, we compute efficiency and effectiveness of those approaches in term of many Deep learning parameters, like MLP, Random forest, RNN, LSTM, that classified instances and time to build model along with other parameters. One major challenge is to extract features from the heartbeat recordings having various noise sources. Different time-frequency and statistical features have been used in automatic heartbeat sound classification.

Project Implementation Method

The structure of this study have three steps: preprocessing, feature extraction and classification model. First we used the preprocessing techniques that remove the noise of heartbeat sound by applying band-pass filter. Afterwards, the sampling frames of heartbeat sound are fixed to 50,000 frame for each heartbeat sound file. After then we used Down-sampling techniques to shorten the sampling frame rate from 50,000 to 782 frames of the heartbeat sound signal. Now we predict the result from the classification model that are applied.

Benefits of the Project

The benefits of this project is to provide the quickest response of classification of heartbeat sound signal with the best accuracy. It will help the doctors to find out the disease.