Abstract
Body mass index (BMI) is most widely used as an indirect measure of fat mass due to its convenience. However, according to previous studies, the poor degree of agreement in obesity classification between BMI and percent body fat was found (Al-Mohaimeed, Ahmed, Dandash, Ismail, & Saquib, 2015; Flegal et al., 2009; Freedman & Sherry, 2009); usefulness of the obesity classification referring to the predefined BMI cut-off value has been controversial. For this reason, this study aimed to find a new method for obesity classification by using four artificial intelligence (AI) techniques and two statistical methods for classification. To enhance its applicability and convenience, we tried to classify obesity with a minimum number of required input body dimensions. The methods were discriminant analysis (DA), kth-nearest-neighbor (kNN), decision tree (DT), support vector machine (SVM), neural networks (NN), and fuzzy rule-based systems (FRBS). FRBS which is an accurate and interpretable classification method can deal with uncertainty and imprecision of the reasoning process, unlike other methods. The performance of the classification methods was achieved in the aspect of accuracy and interpretability. In this study, accuracy refers to the ratio of correct predictions from all predictions made by a model and interpretability refers to the ease of interpretation and application of classification rules. We tried to consider a trade-off between accuracy and interpretability, so that the result could be practically applied to obesity classification. In Korea, females whose body fat percentage are greater than 30% diagnosed as obesity (W. K. Kim, 2010). We defined obesity according to this value. Dataset was obtained from SizeKorea (SizeKorea, 2010). The dataset was collected from 3,224 females aged between 15 and 69 and included variables such as age, height, weight, body fat percentage and BMI. After outliers or omitted cases were eliminated, we analyzed data of which size is 3,180. All possible combinations of variables, which were BMI, height, weight, were used as predicting variable set for each experiment. All analysis conducted in this paper were implemented by R. According to the result, the accuracy of all methods was ranged from 58.77% to 80.22%. Age, in addition to BMI and height/weight, did not give a dramatic effect to improve accuracy; height or weight combined by BMI showed the highest accuracy, in average. DA, DT and SVM accurately classified for all combinations of variables. Except the variable set composed with height and weight, FRBS showed a good performance in the aspect of accuracy, as well. The kNN and NN methods were less precise than the methods previously mentioned. Nonetheless, all methods performed better than the conventional method with the BMI cut-off value. In the case of result with BMI and weight, the accuracy of all methods excluding kNN was more than 10% higher than the accuracy of the method using BMI cut-off values. Among the methods, DA, DT, and FRBS generated classification rules, but in different forms. Rules from DA for classification was in a form of linear equation of two variables, and rules from DT was in a form of tree, which included root, internal nodes and branches. FRBS produced a set of rules expressed in the form of “If A and/or B, then C”, where A, B and C were fuzzy sets. An example of FRBS rule is that “if one person has BMI, which is greater than 25.08 and weight greater than 63.30kg, then she is obese”. DA and DT showed different results if the analysis were iterated. When we compared trials of FRBS, the number of rules was different, but the result became almost identical after eliminating the non-dominant rules. From this, it could be inferred that the result from FRBS was most stable and consistent. Therefore, we suggest that fuzzy rule-based system (FRBS) be the most appropriate method. FRBS performed as accurate as other AI algorithms and DA, and the method provided more stable and consistent classification rules than others. Therefore, this study suggests FRBS using BMI and height or weight for classifying obesity. However, this study has a limitation regarding the level of obese. In a future study, we can subdivide the classes of obesity and conduct the same analysis to investigate methods that are more appropriate. In addition, the methodology can be applied to the male population.