Pathological infant cry analysis using wavelet packet transform and probabilistic neural network M. Hariharan ⇑ , Sazali Yaacob, Saidatul Ardeenaawatie Awang School of Mechatronic Engineering, Universiti Malaysia Perlis (UniMAP), 02600 Perlis, Malaysia article info Keywords: Acoustic analysis Infant cry Wavelet packet transform Probabilistic neural network abstract A new approach has been presented based on the wavelet packet transform and probabilistic neural net- work (PNN) for the analysis of infant cry signals. Feature extraction and development of classification algorithms play important role in the area of automatic analysis of infant cry signals. Infant cry signals are decomposed into five levels using wavelet packet transform. Energy and entropy measures are extracted at every level of decomposition and they are used as features to quantify the infant cry signals. A PNN is developed to classify the infant cry signals into normal and pathological and trained with dif- ferent spread factor or smoothing parameter to obtain better classification accuracy. The experimental results demonstrate that the proposed features and classification algorithms give very promising classi- fication accuracy of 99% and it proves that the proposed method can be used to help medical profession- als for diagnosing pathological status of an infant from cry signals. Ó 2011 Elsevier Ltd. All rights reserved. 1. Introduction Crying is the only way of communication for an infant. From the cry, a trained professional can understand the physical or psycho- logical status of the baby. Infants cry due to some possible reasons such as, hunger, pain, sleepiness, discomfort, feeling too hot or too cold, and too much noise or light. Acoustic analysis of infant cry signal is a non-invasive and has been proven tool for the detection of certain pathological conditions (García & García, 2003; Orozco & García, 2003; Reyes-Galaviz & Reyes-Garcia, 2004; Reyes-Galaviz, Verduzco, Arch-Tirado, & Reyes-García, 2005). In the recent years, simple techniques have been proposed for analyzing the infant cry through linear prediction coding, Mel frequency cepstral coef- ficients and pitch information (García & García, 2003; Orozco & García, 2003; Reyes-Galaviz & Reyes-Garcia, 2004; Reyes-Galaviz et al., 2005). Little attention has been paid by the researchers based on wavelet and wavelet packet transform. This paper presents the development of an intelligent classification system for classifying normal and pathological cry using wavelet packet transform (WPT) and probabilistic neural network (PNN). The application of wavelet and wavelet packet transform analysis is diversified and has been used in many signal and image processing applications. Avci and Avci have proposed a novel approach for radio signal classification based on wavelet packet energy and multi-class support vector machine (Avci & Avci, 2008). Xian and Zeng have proposed an intelligent fault diagnosing method of rotation machinery based on the wavelet packet analysis and hybrid support vector machines (Xian & Zeng, 2009). Wu and Liu have proposed a fault diagnosis system for internal combustion engines using wavelet packet transform (WPT) and artificial neural network (ANN) techniques (Wu & Liu, 2009). Wu and Lin have con- ducted an investigation on speaker identification based on discrete wavelet packet transform with irregular decomposition (Wu & Lin, 2009). Hanbay et al. have proposed a method for the prediction of wastewater treatment plant performance based on wavelet packet decomposition and neural networks (Hanbay, Turkoglu, & Demir, 2008). 1.1. Previous works This section deals with some of the significant works on infant cry signal analysis. Reyes-Galaviz et al. have presented the development of an automatic infant cry recognizer for the early identification of pathologies with the objective of classifying three classes, normal, hypo acoustics and asphyxia (Reyes-Galaviz et al., 2005). They used Mel frequency cepstral coefficients (MFCCs) for feature extraction and a Feed Forward Input Delay neural network with training based on Gradient Descent with Adaptive Back- propagation for classification. The accuracy of their proposed system varies from 96.08% to 97.39%. Orozco and García have developed a method based on linear prediction technique and scaled conjugate gradient neural networks for the detection of pathologies from infant cry. The classification accuracy of their proposed method was 91.08% for 314 samples and 86.20% for 0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.06.025 ⇑ Corresponding author. Tel.: +60 49798419; fax: +60 49885167. E-mail addresses: hari@unimap.edu.my (M. Hariharan), s.yaacob@unimap.edu. my (S. Yaacob), saidatul@unimap.edu.my (S.A. Awang). Expert Systems with Applications 38 (2011) 15377–15382 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa