A. Zabidi

Classification of Infant Cries with Asphyxia Using Multilayer Perceptron Neural Network

Asphyxia occurs in infants with neurological level disturbance, which is found to affect sound of cry produced by infants. The infant cry signals with asphyxia have distinct patterns which can be recognized with pattern classifiers such as Artificial Neural Network (ANN). This study investigates the performance of the Multilayer Perceptron (MLP) classifier in discriminating between healthy and infants with asphyxia from their cries, of ages from zero to seven months old, with an input feature reduction algorithm, Orthogonal Lest Square (OLS) analysis, in contrast to direct selection. The infant cry waveform served as input to Mel Frequency Cepstrum (MFC) analysis for feature extraction. The MLP classifier performance was examined with different combination in number of coefficients, filter bank and hidden nodes. It is found that the OLS algorithm is effective in enhancing the accuracy of MLP classifier while reducing the computation load. Both the average and highest MLP classification accuracies with coefficients being ranked by OLS algorithm have consistently displayed better score than those by direct selection. The highest MLP classification accuracy of 94% is obtained with 40 filter banks, 12 highly ranked MFC coefficients and 15 hidden nodes.

Comparison between Levenberg-Marquardt and Scaled Conjugate Gradient training algorithms for Breast Cancer Diagnosis using MLP

Computerized diagnostic tools have received significant attention over the past few decades, in order to assist medical practitioners in diagnosis of disease based on a variety of test results. It provides a fast and accurate method for diagnosis, particularly in cases where medical practitioners need to deal with difficult diagnosis problems. In this paper, we present an examination of two popular training algorithms (Levenberg-Marquardt and Scaled Conjugate Gradient) for Multilayer Perceptron (MLP) diagnosis of breast cancer tissues. We test the performance of the training algorithms using features extracted from the Wisconsin Breast Cancer Database (WBCD), a benchmark dataset that has been extensively used in literature for breast cancer diagnosis. Based on our results, we conclude that both algorithms were comparable in terms of accuracy and speed. However, the LM algorithm has shown slightly better advantage in terms of accuracy (as evidenced in the average training accuracy and MSE) and speed (as evidenced in the average training iterations) on the best MLP structure (with 10 hidden units).

Performance of Combined Support Vector Machine and Principal Component Analysis in recognizing infant cry with asphyxia

Combined Support Vector Machine (SVM) and Principal Component Analysis (PCA) was used to recognize the infant cries with asphyxia. SVM classifier based on features selected by the PCA was trained to differentiate between pathological and healthy cries. The PCA was applied to reduce dimensionality of the vectors that serve as inputs to the SVM. The performance of the SVM utilizing linear and RBF kernel was examined. Experimental results showed that SVM with RBF kernel yields good performance. The classification accuracy in classifying infant cry with asphyxia using the SVM-PCA is 95.86%.

Binary Particle Swarm Optimization for selection of features in the recognition of infants cries with asphyxia

The infant cry signals with asphyxia have distinct patterns which can be recognized using pattern classifiers such as Artificial Neural Network (ANN). This study investigates the effect of selecting infant cry features using the Binary Particle Swarm Optimization on the performance of Multilayer Perceptron (MLP) classifier in discriminating between healthy and infants with asphyxia from cry signals. The feature extraction process was performed by MFCC analysis. The MLP classifier performance was examined using various combination of number of coefficients. It was found that the BPSO helps to enhance the classification accuracy of MLP classifier while reducing the computational load. The highest MLP classification accuracy achieved was 95.07%, which was obtained when 26 MFCC filter banks, 14 selected MFC coefficients and 5 hidden nodes were used.

Logical Effort using a Novel Discrete Particle Swarm Optimization algorithm

Designers often need to choose the gate sizes for logic circuit designs to estimate the delay of the circuit. Simulation and timing analysis are poor tools for this task because they cannot be modified for better results. Hence, the method of Logical Effort (LE) provides a simple method to overcome these design problems. We apply a novel Discrete Particle Swarm Optimization algorithm (DPSO) to solve the LE problem for electronic circuits. The method uses the rescaling function commonly used to scale datasets in neural networks to convert continuous-valued PSO values into discrete values. The proposed algorithm was successfully applied on a three-stage NAND gate circuit, and has been shown to work well, with 100% accuracy in all test runs.

Heat exchanger modeling using NARX model with binary PSO-based structure selection method

This paper explores the application of Non-Linear Autoregressive Model with Exogenous Inputs (NARX) system identification of heat exchanger system. Model structure selection was performed using the Binary Particle Swarm Optimization (BPSO) algorithm. The application of BPSO for model structure selection represents each particles position as binary values, which were used to select a set of regressors from the regressor matrix. Parameter estimation was then performed using Householder-based QR factorization method. Tests performed on the heat exchanger system defined the model with a maximum lag of five, while fulfilling all model validation criterions.

Classification of infant cries with hypothyroidism using Multilayer Perceptron neural network

Hypothyroidism occurs in infants with insufficient production of hormones by the thyroid gland. The cry signals of babies with hypothyroidism have distinct patterns which can be recognized with pattern classifiers such as Multilayer Perceptron (MLP) artificial neural network. This study investigates the performance of the MLP in discriminating between healthy infants and infants suffering from hypothyroidism based on their cries. The infant cries were first divided into one second segments, and important features were extracted using Mel Frequency Cepstrum Coefficient (MFCC) analysis. Two methods were then used to select which MFCC coefficients to be used as features for the MLP: direct selection or Fishers Ratio analysis (F-ratio analysis). Their performances were compared with experimental results showing that MLP was able to accurately distinguish between the two cases. The classification performance of MLP trained with F-Ratio analysis is found to be better compared to direct selection method.

Non-destructive classification of watermelon ripeness using Mel-Frequency Cepstrum Coefficients and Multilayer Perceptrons

We present a non-destructive watermelon classification method using Mel-Frequency Cepstrum Coefficients (MFCC) and Multi-Layer Perceptron (MLP) neural network. Acoustic signals were collected from thumping noises of ripe and unripe watermelon fruits. MFCC was then used to convert the signals into MFCC coefficients. The coefficients were then used to train a MLP, and the MLP gives the final decision on the watermelon ripeness state. In our paper, we describe the methods used to obtain the acoustic samples, as well as the evaluation of several MLP structures and parameters to obtain the best MLP classifier. Our results show that the proposed method was able to discriminate between ripe and unripe watermelons with 77.25% accuracy.

Optimized Support Vector Machine for classifying infant cries with asphyxia using Orthogonal Least Square

This paper investigates the effect of optimizing Support Vector Machine, with linear and RBF kernels, on its performance in classifying asphyxiated infant cries, with Orthogonal Least Square. Mel Frequency Cepstrum analysis first extracts feature from the infant cry signals. The extracted features are then ranked in accordance to its error reduction ratio with OLS. SVM with linear and RBF kernel then classify the asphyxiated infant cry from the optimized and non-optimized input feature vector. The classification accuracy and support vector number are used to gauge the performance. Experimental result shows that for both kernels, the OLS-optimized SVM achieve equally high classification accuracy with lower support vector number than the non-optimized one. It is also found that the OLS-SVM with RBF kernel outperformed all other methods with classification accuracy of 93.16% and support vector number of 266.2.

Investigation of Mel Frequency Cepstrum Coefficients parameters for classification of infant cries with hypothyroidism using MLP classifier

Hypothyroidism in infants is caused by insufficient production of hormones by the thyroid gland. Due to stress in the chest cavity due to the enlarged liver, the cry signals are unique and can be distinguished from healthy infant cries. We investigate the usage of the Multilayer Perceptron (MLP) classifier to diagnose infant hypothyroidism. The Mel Frequency Cepstrum Coefficients (MFCC) feature extraction method was used to extract important information from the cry signal itself. This study investigates the number of filter banks and coefficients in MFCC to extract optimal information from infant cry signals, to be classified using MLP. The cry signals were first divided into equal-length segments, and MFCC was used to extract features from them. Tests on the combined University of Milano-Bicocca and Instituto Nacional de Astrofisica datasets yielded MLP classification accuracy of 89.18%, suggesting that the optimal MFCC resolution was obtained using 36 filter banks, and 19 coefficients.