2010 IEEE EMBS Conference on Biomedical Engineering & Sciences (IECBES 2010), Kuala Lumpur, Malaysia, 30th November - 2nd December 2010. Detection of Asphyxiated Infant Cry using Support Vector Machine Integrated with Principal Component Analysis R. Sahak i , Y. K. Lee 2 , W. Mansor, A. 1. M. Yassin, A. Zabidi Facult of Electrical Engineering Universiti Teknologi Mara, 40450 Shah Alam, Selangor, Malaysia 'rohilahsahak@yahoo.com 'leeyootkhuan@salam.uitm.edu.my Abstract-Asphyxia refers to respiratory failure in infants, a condition caused by inadequate intake of oxygen. It is important to diagnose asphyxia in infants as early as possible, as it could lead to infant morbidity. PCA has the capability to reduce the dimension of input feature vector to SVM. Previous attempts with PCA and SVM to detect asphyxia from baby cries found their principal components in a random manner, which consumes tremendous computation effort and time. Our work here investigates the improvement in performance to detect asphyxia from infant cries by integrating PCA and SVM with a polynomial kernel, with principal components being ranked by EOC, CPV and SCREE methods. Extracted features from the analysis of MFC coefcients are first ranked with the three feature selection algorithms of PCA, before being submitted to SVM for classifcation. Classifcation accuracy and support vector are employed to gauge the performance. It is found that the highest classification accuracy and support vector number from classifcation with support vector machine alone are 93.836% and 335.1, with a second order polynomial kernel and a regularization parameter of 1 E-04, while those from CPV and SVM outperformed with CA of 94.172%, a low SV of 254.3, a third order polynomial and regularization parameter of 1 E-05. Keywords- Asphyxia, infant cry, principal component analysis, support vector machine, polynomial kernel I. [NTRODUCTION Asphyxia refers to respiratory failure in infants, a condition caused by inadequate intake of oxygen [I]. [t is important to diagnose asphyxia in infants at early birth since it is one of the causes of infant morbidity. [f improper treatment is given to the infant, hypoxia can happen due to insufcient supply of oxygen to the brain, organ and tissues. This will lead to further complications that would result in death. Previous studies found relationship between infant cry and asphyxiated infants [2-3]. A classifcation accuracy of 86% in the detection of asphyxia and deafess was attained with feed forward input delay neural network [4]. The classifcation accuracy is found to have improved to 97.4% with the addition of principal component analysis (PCA) to reduce the dimension of feature vector [5]. Then, support vector machine (SYM) and fuzzy support vector machine (FSYM) were applied to the same classifcation problem, with PCA for preliminary feature extraction. It is found that FSYM with 10 principal components outperformed SYM to reach a classifcation accuracy of 94.982% [6]. All the studies above made use of the transformed coeffcients of infant cry signals from mel fequency ceptrm (MFC) analysis. However, the number of principal components for use by the PCA is selected randomly and hence higher classifcation accuracy is possible. Besides, this process consumes tremendous computation time and effort. Our work introduces and compares three mathematical approaches to selection of principal components of signifcance, i.e. eigenvalue-one-criterion (EOC), cumulative percent variance (CPY) and scree test (SCREE), to make selection of principal components in a more mathematic and systematic way, to allow more convincing performance evaluation at lesser computation time. SYM with polynomial kernel is used in the fnal stage to classif the cry patterns. Classifcation accuracy and number of support vector are adopted as the comparison platform. It is hope that automated detection of asphyxiated infants fom their cries, once validated, could help to provide relief to nursing staff in screening, more timely treatment to needy infants that would lessen casualties, less calculated risk with more true positives and true negatives due to fatigue of medical staff. II. FEATURE SELECTION AND PATTERN CLASSIFICATION Support vector machine is used to classif normal infants fom the asphyxiated ones, tracing their cries while principal component analysis is applied to select feature of signifcance from the transformed coeffcients. The uniqueness of the integration of these two techniques to our work being optimisation of principal components to reduce the dimension of input feature vector, which comes with the advantage faster solution at lesser computation effort and time. The following gives a summarised background theory on the two techniques. A. Principal Component Analysis(PCA) Redundant parameter is a common phenomenon of feature extraction, which results in the presence of features with minimal signifcance and unnecessary large dimension. PCA has the advantage of fi ltering for the signifcant features while fltering off the redundant features. It entails a mathematic procedure that transforms variables into employment variables known as principal components (PC). The transform carries three obvious effects; it orthogonalizes the components of the 978-1-4244-7600-8/10/$26.00 ©2010 IEEE 485