Detection of infant hypothyroidism with mel frequency cepstrum analysis and multi-layer perceptron classification

Hypothyroidism in infants is caused by insufficient production of hormones by the thyroid gland. Due to stress in the chest cavity as a result of the enlarged liver, their cry signals are unique and can be distinguished from healthy infant cries. Our work investigates the effectiveness of using Multilayer Perceptron classifier to detect infant hypothyroidism. The Mel Frequency Cepstrum coefficients feature extraction method was used to extract vital information from the cry signals. The number of hidden units and MFC coefficients for optimal performance were also investigated. The cry signals were first divided into equal length segments of one second each and MFC analysis was performed to produce the coefficients as input feature vector to the MLP classifier. Tests on the combined datasets from University of Milano-Bicocca and Instituto Nacional de Astrofisica yielded MLP classification accuracy of 88.12%, area under curve of 99.89%, with 15 hidden units and 20 coefficients, being the most optimal MFCC resolution.