Non-linear features and feature selection algorithms for speech based prediction of body mass index (BMI)

Abstract

Obesity and overweight have been a growing concern due to their negative impacts on human‘s health. Obesity is considered as a major cause of some serious diseases such as diabetes, cardiovascular diseases, and metabolic syndrome, and it has become epidemic. Today, body mass index (BMI) is widely used as a tool to classify normal weight, overweight, underweight and obesity. These measurements are sometimes not suitable for remote healthcare or u-healthcare supporting general treatment and emergency medical service in real time at remote locations. The researchers have explored the association between speech recognition and BMI. Speech signals have a close relation with BMI status, which is predicted by a combination of key features. The purpose of this research work is to predict BMI status (normal, overweight and obese) using speech signal without weight and height measurements. In this research work, wavelet packet based nonlinear entropy features and feature selection algorithms were proposed to predict BMI status via speech signal of normal, obese and overweight subjects. The recorded speech signal (/ah/ sounds) were decomposed up to level five using wavelet packet transform (WPT). Several features were extracted from the wavelet packet coefficients and an Analysis of Variance (ANOVA) test.