Application of Neural Networks in the Classification of Obesity Using Deurenberg Equation

Direct methods for accurately measuring the volume of fat tissue in the human body are not suitable for use in epidemiological studies due to their lack of precision. Therefore, indirect methods, such as indices and formulas, are used to estimate the percentage of body fat (BF%) in patients. While the body mass index (BMI) is a commonly used method, it does not differentiate between adipose tissue and other tissues, and only provides a height-weight ratio. Recent studies have shown that machine learning algorithms, such as support vector machines (SVMs), can effectively evaluate anthropometric parameters to obtain BF%. Although the use of artificial neural networks (ANNs) for this purpose remains a possibility for future research. The objective of this research is to evaluate the effectiveness of ANNs in classifying abnormal values of BF%. The ANNs were trained using a database consisting of 7750 subjects, ranging in age from 17 to 102 years, and including 27 predictive variables. To evaluate the ANNs the accuracy (ACC), specificity (SPE), sensibility (SEN), positive predictive value (PPV), negative predictive value (NPV) and F1 were calculated. The results suggest that ANNs are more effective than SVMs in addressing classification impaired BF%, although the difference between the two methods is not significant when compared to previous work that used the same database. Notably, ANNs outperformed SVMs in their ability to classify true negative values, achieving a success rate of NPV of above 98% compared to 85% for SVMs. This finding is consistent with the results obtained for specificity, where ANNs achieved a rate of above 99% compared to 67% for SVMs. Additionally, all other metrics (ACC, SEN, SPE, PPV, and F1) yielded results above to 96%.