Discrete Mutative Particle Swarm Optimisation of MFCC computation for classifying hypothyroidal infant cry

摘要

This paper describes the optimization of Mel Frequency Cepstral Coefficients (MFCC) parameters using Discrete Mutative Particle Swarm Optimization (DMPSO) for diagnosis of hypothyroidism in infants. The MFCC was used to extract the feature set from infant cry signals. The features were then classified using Multi-Layer Perceptron (MLP). The DMPSO variants optimize the number of filter banks and number of cepstral coefficients in MFCC. Based on the values chosen by DMPSO, the extracted features were then fed to 50 MLP classifiers (with different initial weight initialization values), which were trained to discriminate between healthy and hypothyroid infants. The results showed that DMPSO managed to produce classification accuracy of 88.7% with percentage convergence of 66.7% in detecting hypothyroidism from infant cry signals. The optimal number of filter bank and MFC coefficients were found to be 36 and 19 respectively.