Detection of asphyxia from infant cry by linear kernel support vector machine enhanced with features from orthogonal least square

Abstract

An investigation into the performance of SVM with linear kernel and features ranked by OLS, to discriminate infants with asphyxia from their cries, is presented in this paper. The features of the cry signal were first transformed into MFC coefficients. The input feature set was then used for classification by SVM with linear kernel. The number of coefficients and filter banks were tuned to acquire the optimal input feature set. This is uniquely different from previous works, where empirical values were simply adopted without proof. However, it is found that the performance of the classifier can be improved further by using selective coefficients from the optimal feature set. Hence, the MFC feature coefficients were then ranked in accordance to its error reduction ratio using OLS before submission to the classification stage. From experimental works, it was found that the optimal input feature set for DS-SVM approach is obtained with 20 coefficients, 21 filter banks and regularization parameter of 0.001 while the OLS-SVM approach reduced the MFC coefficients to 14. From performance comparison of both, it can be concluded that the OLS-SVM excelled the DS-SVM approach at classifying infant cry with asphyxia. This is because the OLS-SVM approach yields comparable classification accuracy (92.5%) with lesser support vector number (252.5) and lesser MFC coefficients (14) than the DS-SVM approach, which implicates much reduced computation effort and load.