Diagnosis Of Childhood Pneumoniaoutside The Hospital Based On Coughsound

Pneumonia annually kills around 1 million 400 thousand children throughout the world,of which 75%occurs in remote areas of poverty in Africa and Asia.Prompt diagnosis and effective treatment can prevent a large number of pneumonia deaths.However,the lack of professional medical equipment and team in these areas is difficult to get accurate pneumonia diagnosis in time.In addition,children’s hospitals in big cities are often overcrowded in the flu season.Most of them are only common cold.They can rest at home and do not need to go to the hospital immediately.In this paper,using the speech recognition technology,we proposed a method of pneumonia recognition based on cough sounds to realize the diagnosis of children pneumonia outside the hospital.The purpose of this method is to distinguish the pneumonia cough sound from other pathological cough sound,recognition process like speech recognition.Firstly preprocess cough sound signal,and then extract feature,finally input feature vector to the classification model which has been trained to predict result.The main work of this article is as follows:1)Improved the standard MFCC feature vector,MFCC extraction process using Mel frequency filters group which distribution in the frequency domain for the performance of low frequency dense and high frequency sparse.But pneumonia and other pathological cough sound difference in the frequency domain is mainly reflected in the medium band,so MFCC cannot effectively to reflect the two differences between the cough sound.Therefore,the distribution of the Mel frequency filter bank is restructured to be dense to medium frequency,low frequency and high frequency sparsity.The experimental results show that the improved 16order MFCC1 has better classification performance than the standard 16 order MFCC,and the accuracy rate rises from 55.78%to 59.38%.2)Proposed a new feature extraction method,wavelet energy cepstrum feature.The difference of two kinds of cough sound signals in the frequency domain can be reflected as the difference corresponding to the energy distribution.Therefore,use wavelet packet transform to get the energy coefficient of each band,then keep the biggest difference of two kinds of signal band by variance analysis,finally cepstrum the feature vector to make it more effective.The classification accuracy of the feature is 61.7%.3)Use SVM model classifier to recognize whether a single cough is pneumonia or not.By comparing the effects of different kernel functions,parameter values and feature vector combinations on the experimental results,the final choice of the kernel function is radial basis kernel and parameter values isg（28）08.、C（28）1.Feature vector is MFCC1 and its first-order difference,wavelet energy cepstrum,short-time energy coefficient and children information signs.The accuracy of classification is 73.26%.4)On the basis of the recognition of a single cough,the children were identified as pneumonia or not through the identification of multiple cough sounds.The ROC analysis curve was used to obtain the PCI_thh value with the highest overall accuracy,so as to make the sensitivity and specificity of the pneumonia identification results relatively high.The final PCI_thh value is set to 0.6,the sensitivity is 91.3%,the specificity is 92.9%,and the overall accuracy rate is 91.9%.5)A software system based on the above algorithm is developed,which provides users with an environment of friendly GUI.This topic not only has theoretical significance,but also has practical significance in engineering.