Using an improved relative error support vector machine for body fat prediction

Abstract

<p><b>Background and Objective:</b> The term 'obesity' refers to excessive body fat, and it is a chronic disease associated with various complications. Although a range of techniques for body fat estimation have been developed to assess obesity, they are typically associated with high-cost tests requiring special equipment. Accurate prediction of the body fat percentage based on easily accessed body measurements is thus important for assessing obesity and its related diseases. This paper presents an improved relative error support vector machine approach to predict body fat in a cost-effective manner.</p> <p><b>Methods:</b> Our proposed method introduces a bias error control term into its objective function to obtain an unbiased estimation. Feature selection is also utilised, by removing either redundant or irrelevant features without incurring much loss of information, to further improve the prediction accuracy. In addition, the Wilcoxon rank-sum test is used to validate if the performance of our proposed method is significantly better than other prediction models being compared.</p> <p><b>Results:</b> Experimental results based on four evaluation metrics show that the proposed method is able to outperform other prediction models under comparison. Considering the characteristics of different features (e.g., body measurements), we show that applying feature selection can further improve the prediction performance. Statistical analysis carried out confirms that our proposed method has obtained significantly better results than other compared methods.</p> <p><b>Conclusions:</b> We have proposed a new approach to predict the body fat percentage effectively. This approach can provide a good reference for people to know their body fat percentage with easily accessed measurements. Statistical test results based on the Wilcoxon rank-sum test not only show that our proposed method has significantly better performance than other prediction models being compared, but also confirm the usefulness of incorporating feature selection into the proposed method.</p>