Predicting body fat percentage based on gender, age and BMI by using artificial neural networks
Introduction
The prevalence of obesity has increased dramatically worldwide during the past few decades. Obesity is recognized as an independent factor for the development of the cardiovascular diseases that are among the main causes of death in the world [1]. Among the main cardiovascular risk factors, the obesity is of special importance since the increase of fat mass launches a cascade of adipokine-mediated metabolic, inflammatory and homeostatic disturbances accelerating the process of atherosclerosis [2], [3], [4]. The most intensive changes take place inside the visceral fat depots and that is the reason why the abdominal or central type of obesity is marked as particularly risky [5], [6]. Obesity implies increased body weight due to the enlargement of the adipose tissue to the extent that impairs health. Generally, the diagnosis of obesity depends on total body mass (BM), body height (BH) and body fat mass, with fat distribution.
The body fat percentage (BF%) is the measure of total body fat and an accurate measurement is necessary for the diagnosis of obesity. There are numerous techniques for estimating BF%, such as: anthropometry (body mass, circumferences and diameters, skinfold thickness etc.), underwater weighing (UWW), dual energy X-ray absorptiometry (DEXA), bioelectrical impedance analysis (BIA), computed tomography (CT), magnetic resonance imaging (MRI) and near infrared interactance [7], [8], [9].
Anthropometric measurement is the most common method and also the cheapest for the estimation of BF%. Body mass index (BMI) has been widely accepted as a simple and the most practical measure of obesity in clinical and epidemiological studies, eventhough it does not distinguish fat from lean body mass. BMI is calculated as the ratio of BM and the square of BH. The values of BMI over 25 kg/m2 correspond to the overweight, and values over 30 kg/m2 correspond to obesity [10]. BMI provides a measure that allows the comparison of the adiposity of individuals of different heights and weights, but does not provide sufficient information about fat mass. In fact, it is an indicator of the nutritional status, not a measure of body fat mass. The process of aging is characterized by decreasing body height and increasing fat mass with redistribution of fat tissue by means of visceral fat deposition, even if body weight and BMI are maintained [11], [12]. On the other hand, during the entire adult life span the BF% of females is significantly higher than that of males with the same BMI [13], [14]. Based on that, the relationship between BF% and BMI is gender- and age-dependent [15], [8]. There are numerous formulas (Deurenberg et al. [16], [17], Gallagher et al. [13], Jackson and Pollock [18], [19], Jackson et al. [14] etc.) for estimating BF% from GEN, AGE and BMI. These linear formulas are all fairly similar and use coefficients that are previously determinated by using statistical methods. However, there are uncertainties and controversies whether the relationship between BMI and BF% is linear or curvilinear [8].
The DEXA and UWW are more accurate than anthropometric measurements and widely acceptable as gold standards in epidemiological studies, but their cost and complexity limit the use [7]. The BIA is a lower-cost and more suitable method with respect to DEXA and UWW. In this study, body fat percentage BF% was measured using Tanita bioelectrical impedance analyser (Tanita-BIA). Tanita-BIA provide good agreement with DEXA and valid measures of body fat percentage, so it could be a convenient and practical approach for assessment in community-based research [9], [20]. Tanita-BIA measurements correlate highly with both the DEXA and UWW, much better than anthropometry [21].
Our goal is to predict BF% by using artificial neural network (ANN) as a well-known method of artificial intelligence. Clearly, in this paper, the feed-forward ANN with back-propagation as the training algorithm has been applied to estimating BF% based on GEN, AGE and BMI. ANN has proven to be a useful predictive tool in numerous clinical fields [22] (e.g. prediction of the heart attack [23], metabolic syndrome [24], [25], level of cardiometabolic risk [26], WHtR and SAD limits [27], [28] etc.). The ANN systems have been extensively used to address complex real-world problems, particularly when dealing with non-linear models or when the definition of underlying mechanisms is incomplete. ANN takes known data, i.e. previously solved examples, recognizes complex patterns between inputs and outputs and then applies this knowledge on unknown data. The hidden relationships between inputs and outputs are learned so subsequently ANN is able to predict the output from a given input of new data [29]. ANN may represent a valid support owing to its ability to find patterns in data involving many variables [30].
In this study, we use MATLAB, Version 7.11.0.584 (R2010b). Section 2 describes our methodology, i.e. the observed parameters and their measurements, the obtained dataset and ANN solution for predicting BF%. The results and contribution of our solution are discussed in Section 3. The paper ends with conclusions.
Section snippets
Measurements
The group inquired consisted of 2755 subjects (1332 women and 1423 men) aged 18 to 88 y, with BMI values between 16.60 and 64.60 kg/m2 and with BF% values between 3.80% and 71.70%.
Table 1 presents the average values and ranges (min.–max.) of the parameters considered. Data is depicted separately for women and men and subsequently for the whole group examined.
All subjects were from the north part of Serbia. The study was conducted in accordance with the Declaration of Helsinki. The subjects
Optimal ANN architecture
We have investigated single hidden layer ANN architectures with Nh = 1, 2, … 40 hidden neurons. Table 6 from Appendix contains average values and standard deviations of the mean square errors in the validation phase and testing phase . These values were obtained in MATLAB respectively from tr . best _ vperf and tr . best _ tperf.
Comparison of the results from Table 6 leads to the conclusion that single hidden layered ANN architecture with 31 hidden neurons has the
Limitations
This research was limited solely to the Serbian population. However, it is known that the relationship between BF% and BMI differs among different ethnic groups [17]. In order to develop a software application for the predicting BF% we need an international database.
Conclusion
This paper presented an ANN solution for predicting BF% with accuracy 80.43%. The solution was compared to Deurenberg et al. [16], [17], Gallagher et al. [13], Jackson and Pollock [18], [19], Jackson et al. [14] formulas. ANN showed higher predictive accuracy for +1.23% to +3.12%. Based on that, we conclude that this paper presented a new approach to predicting BF% that has same complexity and costs but higher predictive accuracy.
Acknowledgements
This work was partially supported by the Ministry of Education, Science and Technological Development of the Republic of Serbia within the projects: ON 174026 and III 044006.
References (36)
- et al.
The synthesis of the rough set model for the better applicability of sagittal abdominal diameter in identifying high risk patients
Comput. Biol. Med.
(2010 Sep) - et al.
Clinical and pathophysiological consequences of abdominal adiposity and abdominal adipose tissue depots
Nutrition
(2003) - et al.
The relationship between BMI and percent body fat, measured by bioelectrical impedance, in a large adult sample is curvilinear and influenced by age and sex
Clin. Nutr.
(2010) - et al.
Waist circumference and abdominal adipose tissue distribution: influence of age and sex
Am. J. Clin. Nutr.
(2005) - et al.
Aging of adipocytes: potential impact of inherent, depot-specific mechanisms
Exp. Gerontol.
(2007) Application of artificial neural networks to clinical medicine
Lancet
(1995)- et al.
Prediction of metabolic syndrome using artificial neural network system based on clinical data including insulin resistance index and serum adiponectin
Comput. Biol. Med.
(2011) - et al.
A primary estimation of the cardiometabolic risk by using artificial neural networks
Comput. Biol. Med.
(2013 Jul) - et al.
Introduction to neural networks
Lancet
(1995) - Global health risks: mortality and burden of disease attributable to selected major risks. Geneva: WHO, December 2009....