A simple equation to estimate body fat percentage in children with overweightness or obesity: a retrospective study

Study population

Patients referred for nutritional problems to the Nutrition, Growth and Metabolism Unit of the Department of Pediatrics of San Juan de Alicante University Hospital. This hospital covers an area of about 220,000 inhabitants in the province of Alicante, which is located in the southeast of Spain and has a total of 1,843,589 inhabitants (Instituto Nacional de Estadítica, 2016). The health system is universal and free for both children and adults. The prevalence of childhood obesity in the province of Alicante is approximately 13.5–18.8% (Ruiz & Pérez et al., 2008). The criterion for referral by the primary care pediatrician was overweight/obesity.

Study design and participants

We conducted a cross-sectional observational study in patients aged 4–18 years who met the following inclusion criteria: having had an initial referred visit to the Nutrition, Growth and Metabolism Unit of San Juan de Alicante University Hospital and a DXA assessment prior to the implementation of dietary measures and lifestyle changes. The exclusion criteria encompassed excess weight secondary to causes other then high caloric nutrition, such as growth hormone deficiency with hormone replacement therapy, syndromic obesity, obesity secondary to other diseases such as hypothyroidism, protracted treatment with corticosteroids or other drugs that could influence energy intake or expenditure, precocious puberty, or neurologic or neuromuscular disorders preventing the patient from walking or exercising normally. The data were collected between July 2007 and July 2016, and all the study children were Caucasian.

Variables and measurements

Our main outcome variables were body fat (% total body fat) and fat mass index (FMI, in kg/m²). These were measured with a General Electric Lunar DPXN PRO™ DXA densitometer (GE Healthcare, Little Chalfont, Buckinghamshire, United Kingdom). The software of this device is able to obtain measurements for the total weight and percentage of body fat. These parameters were used as the reference standard measurements.

As secondary variables we used BMI (in kg/m²), sex (male or female) and age (in years). To obtain the anthropometric measurements (weight and height), we followed a standardized protocol with a stadiometer accurate to 0.1 cm and a SECA scale accurate to 0.1 kg. Each value was measured twice by a single individual, and the mean of the two measurements was used in the analysis.

Sample size calculation

The sample collected during the study period was 416 children. This sample was randomly divided into two equal parts (n = 208). The first group was used to construct a predictive model (multivariable linear regression) and the second group was used to validate it.

Construction: To construct a multiple linear regression model we must look at the relationship between the number of subjects and the number of predictors (subject-to-variable ratio). As a heuristic rule, it was considered that this ratio should be at least 50, which allowed the introduction of four explanatory variables in the predictive model.

Validation: The sample consisted of the other 208 subjects. This sample size was very satisfactory in obtaining excellent predictions, since the constructed models had four predictors and coefficients of determination of 0.65 and 0.86 for fat percentage and FMIrespectively (Knofczynski & Mundfrom, 2008). The coefficient of determination values were obtained in the construction sample and were used to determine the sample size of the validation.

Statistical analysis

Continuous variables (BMI, FMI, age and body fat) were summarized using means and standard deviations. To describe the sex variable, we calculated absolute and relative frequencies. To compare the homogeneity of the construction and validation samples, we performed the t-test and the Pearson’s chi-squared test. We calculated the following exponentiations: bases (BMI and age) and exponents (−2, −1, 1 and 2). All the interactions between sex and the exponentiation variables were obtained. In the construction sample (n = 208) we aimed to construct a linear regression model to predict body fat. Taking into account that we could only introduce 4 explanatory variables (subject-to-variable ratio: one per each fifty subjects) and had 17, we applied the following algorithm in order to select them: we obtained all the possible combinations of 1, 2, 3 and 4 elements from the total (17 variables): 3,213 combinations. In other words, we analyzed 6426 different models (3,213 × 2 ≈ 6,500). In each combination we estimated the linear regression model with them in order to predict the body fat and calculated its coefficient of determination. The combination/model with the maximum coefficient was then selected. The goodness-of-fit of the model was assessed with the ANOVA test. The model was internally validated using bootstrap methodology in the validation sample, calculating the intraclass correlation distribution (two fixed judges (the predicted and observed value) with absolute agreement in the ratings). In addition, the scatter plot with the estimation and the real parameter was obtained. Finally, we also calculated the Bland & Altman limits of agreement (Bland & Altman, 1986). The same process was applied for FMI. We set the level of statistical significance at 0.05. The statistical software used was IBM SPSS Statistics 19 and R 2.13.2.

Ethical considerations

This study adhered to the ethical principles of the Declaration of Helsinki, and only involved the performance of procedures used in everyday clinical practice. The data were processed safeguarding anonymity and confidentiality, and the study was approved by the corresponding Ethics Committee (Comité Ético de Investigación Clínica del Hospital Universitario de San Juan de Alicante, ref 16/305). Informed consent was not requested from the parents for this study, since it was part of routine clinical practice without any type of intervention. The Ethics Committee approved this procedure.

Summary

We constructed and validated two predictive models to determine body fat and FMI in children with overweight and obesity, using the values obtained by DXA as the reference standard. Both models showed excellent accuracy (ICC > 0.75) (Cicchetti, 1994).

Strengths and limitations of the study

The main strength of the study is the easy and cost-effective calculation of body fat percentage and FMI based on the widely used standard anthropometric measurements (height and weight), age and sex. Although more complex techniques, such as DXA, consume little time and few resources and involve a very low radiation exposure, they are not widely available in health care settings and are more complicated and costly to implement. We also underscore the statistical methodology followed for its construction, which took into account the subject-to-variable ratio of about 6,500 models. For validation, the sample size was adequate to obtain excellent predictions.

Regarding limitations, the study was clearly lacking in normal-weight cases (selection bias), but due to ethical and cost-of-care considerations, body fat measurement using DXA was not justified in children of normal weightand was only performed for other clinical purposes, such as in patients that required a bone mineral density assessment. On the other hand, information bias was minimized through the use of fully validated and calibrated devices, and confounding bias was minimized through the estimation of multivariable models. Thus, our models can only be applied in children with overweight or obesity. Furthermore, although DXA is not considered the gold standard for the assessment of body fat because a four-component model is better (Sopher et al., 2004; Sopher, Shen & Peitrobeilli, 2005; Williams et al., 2006), DXA is generally considered to provide the most valid estimates in clinical practice (Ellis, 2001). Finally, the models only included anthropometric parameters; that is they did not consider other factors associated with body fat (Moliner-Urdiales et al., 2009; Van Sluijs et al., 2010; Gómez-Martínez et al., 2012). Nevertheless, even without these other factors our predictions were still precise.

Comparison with the existing literature

When comparing our predictive models with the existing literature, we find studies that evaluate a single anthropometric parameter and other authors who developed multivariable predictive models based on skinfolds (Silva et al., 2013; Jensen, Camargo & Bergamaschi, 2016). We must bear in mind that the studies evaluating a single anthropometric parameter (circumferences of the upper arm, waist, hip and others; ratios such as waist-to-height or waist-to-hip) (Freedman et al., 2004; Freedman et al., 2005; Freedman et al., 2012; Bergman et al., 2011; Goossens et al., 2012; Boeke et al., 2013; Weber et al., 2013; Craig et al., 2014) did not assess the power of this parameter, the interactions with other variables, or a combination of factors to estimate body fat, which gives greater accuracy to the results obtained in our model. Multivariable models of using skinfolds have the clinical drawback of skinfolds being difficult to measure and the methodological drawback that they did not follow the statistical techniques that yield the highest power in the construction and validation of a multiple linear regression predictive model (interactions, powers, bootstrapping, ICC and testing of approximately 6,500 models) (Silva et al., 2013; Jensen, Camargo & Bergamaschi, 2016). In light of the above, our model clearly provides greater accuracy with respect to the others published in the scientific literature.

Implications for research and practice

The diagnosis, treatment and follow-up of obesity in the pediatric age group have become a global health priority. Consequently, there is great interest in the healthcare field in the development of quick and accurate tools that can be used in the follow-up of these patients. In this study, we developed formulas for the calculation of body fat percentage and FMI based on BMI, age and sex, which facilitate monitoring of adiposity in the management of these patients, reserving the use of more accurate methods such as DXA for extreme cases.

Regarding possible future lines of research, we encourage other authors to externally validate the equations developed in this paper. The methodology used in this work can be applied to create new equations for body fat or for other types of parameters, both anthropometric and non-anthropometric.