Research Of Mobile Phone PCG Record Classification Based On Convolutional Neural Network

Cardiovascular disease has always been a big threaten to human health,the death number of which ranks first among all fatal diseases'.Therefore,early diagnosis is important and necessary to the prevention and cure of this kind of disease.As a research target of this article,the phonocardiogram(PCG)is an indicator that reflects the health condition of heart.PCG represents a series of sounds that arise from heartbeat,thus it can provide various information of heart and is helpful in the healthcare of cardiovascular disease.On the other hand,as the technique of artificial intelligence(AI)developing,lots of diagnosis methods that depend on AI technique appear,making it possible to assist manual diagnosis with machine diagnosis.This advancement may satisfy the need of regions that lack medical resources or individual's daily diagnosis.This article aims to implement a portable PCG classification function on mobile phone.We classify heart sound that recorded by mobile phone into healthy or not based on convolutional neural network(CNN).The database used in this research consists of the Physio Net Challenge 2016 database and PASCAL heart sound database.After data cleaning,PCG signals are transformed into two dimensional spectrogram.Then we design a CNN model for classification and it has an accuracy of 85.35%.Furthermore,considering that mobile phone record has some difference with the original PCG data,training with these original data although has a high accuracy but is not appropriate in actual use,so we do some processing on the data to imitate record quality.By training with the new data,the accuracy,sensitivity and specificity of this algorithm is 83.98%,84.62% and 83.37%.In addition,we developed a Wechat Mini Program based on the model,which can do PCG recording and loading,then classify PCG on its server.On a real-world test set containing 20 PCG record,the model trained by original data has an accuracy of 50.0% while the model trained by processed data has an accuracy of 75.0%.