Development and validation of body fat percent prediction equation for screening obesity in adults

SUMMARY

Background:

In Ethiopia, use of advanced body composition measurement methods may not be feasible due cost and unavailability of the facilities. This study developed and validated body fat percent prediction equation for adults using locally appropriate data.

Methods:

The study was conducted from February to April 2015 among 704 randomly selected adult employees of Jimma University. The total sample was spilt and randomly assigned to a training (n = 352) sample used for developing Ethiopian body fat percent (BF%) prediction equation and a testing (validation) sample (n = 352) used for determining the validity of the equation. A multivariable linear regression model was used to develop BF% prediction equation on the training sample using Air displacement Plethysmography (ADP) measured BF% as dependent variable and age, sex and body mass index as predictor variables. For the testing (validation) sample, BF% measured using ADP and the one predicted using the newly developed Ethiopian and Caucasian BF% prediction equations were compared using validity measures, Kappa statistics and agreement between the two measures was determined using Bland Altman plot.

Results:

A multivariable linear regression model run on the testing population showed that age, sex and BMI were significant predictors of ADP measured BF%. Accordingly, the BF% prediction equation of Ethiopian adults was generated as follows: BF% = −8.601 + BMI (1.521) + Age (0.243) + Sex (−10.568), where sex = 1 for males and 0 for females. Comparison of measured and predicted BF% showed that there was no significant (P = 0.932) difference between ADP measured BF% and BF% predicted using Ethiopian equation with a mean (±SD) difference of 0.03 (±5.44). Conversely, there was a significant difference (<0.0001) between ADP measured BF% and the Caucasian Equation estimated BF% with a mean (±SD) of 6.83 (±5.57). In both males and females, the Ethiopian equation demonstrated a very good to excellent sensitivity, specificity, positive predictive value and negative predictive values. Conversely, the Caucasian equation had poor sensitivity and negative predictive values, while it demonstrated an excellent specificity and positive predictive value. Likewise, there was a substantial Kappa agreement for males (K = 0.741) and for females (K = 0.720) between Ethiopian equation and ADP in diagnosing obesity among males based on BF%, while there was a slight Kappa agreement for males (K = 0.156) and a fair Kappa agreement for females (K = 0.365) between Caucasian equation and ADP (P < 0.001).

Bland Altman plot showed a good agreement between ADP measured BF% for the Ethiopian Equation and not for the Caucasian equation. It was observed that the Ethiopian equation has a better prediction of BF% when compared to the measured one, but the Caucasian equation consistently underestimated BF% for all samples with different levels BF%.

Conclusion:

The new Ethiopian BF% prediction equation performed very well in predicting BF% in the testing population in terms of validity measures, Kappa agreement and Bland Altman plot; while the Caucasian equation significantly underestimated body fat percent among Ethiopian adults. The results imply that the new Ethiopian equation can be used as a cost effective and user friendly screening method for early detection of obesity for the prevention of associated morbidity and mortality in Ethiopian adults.

1. Background

Accurate measurement of body fat percent (BF %) can be used to evaluate the effect of exercise and dietary interventions in weight loss programs [1]. Body Mass Index (BMI) is the most widely used simple tool to identify the problem of obesity and risk of metabolic diseases [2]. Studies from developed countries indicated that there are ethnic and racial differences in the capability of BMI in predicting BF% [3–6]. BMI cannot differentiate two individuals with the same weight but different BF% [7–9].

This makes the utility of international BMI cut-off for the determination of body composition among different populations questionable. A meta-analysis showed that international BMI cutoff failed to identify half of individuals with excess BF% [10]. As there are variations of body fat distribution among different ethnic groups, BF% measurement allows for more precise estimation of body composition than BMI [11].

BF% is recommended to be powerful indicator of obesity and risk of chronic non-communicable diseases [10–18]. Although non-communicable diseases (NCDs) associated with increased body fat percent have been problems of affluent countries; their magnitude is increasing in developing countries since the past few decades [19–21]. According to WHO, the prevalence of chronic non-communicable diseases exceeds that of communicable diseases in sub-Saharan African countries including Ethiopia by 2030 [20]. In Ethiopia, the prevalence of cardiovascular diseases (CVD) increased dramatically in the past few years with 31.5% of men and 28.9% of women in Addis Ababa having high blood pressure, indicating a “silent epidemic” of CVDs [22] accounting for over a third of deaths [23]. The problem is expected to take rather a sharp turn to the worst in Ethiopia due to high prevalence of childhood stunting, which is hypothesized to enhance risk of NCDs due to associated organ stunting [24].

Monitoring and vigilance of such epidemiologic transmission is required to prevent the double burden of both chronic non-communicable diseases and infectious diseases [25]. Although there are several advanced methods of measuring body fat percent, they are too expensive for routine service use, require qualified personnel and are not portable for use at the community level [2,26].

Body fat percent prediction equation has been developed from body mass index based on BF% measured using advanced techniques [17]. Studies showed that the relationship between BMI and body fat percent varies based on ethnic differences [27] making BF% prediction equations developed for one population inappropriate for the another. A meta-analysis of studies on the relationship between measured body BF% and predicted BF% using Caucasian equation showed a significant underestimation of BF% among Ethiopians, Thais and Polynesians [5], which was hypothesized to be due to the differences in body build. However, this finding was based on a very small sample of Ethiopians, which makes the estimates less reliable. Moreover, the performance of the equation was not validated using a testing (validation) sample. Therefore, the current study aimed to develop BF% prediction equation for Ethiopian adults and validate it using Kappa statistics, validity measures and agreement with ADP measured BF% using Bland Altman Plot.

2. Materials and methods

The study was conducted from February to March, 2015 in Jimma University (JU), located 357 km southwest of Addis Ababa. The university has two institutes and six colleges housing a total of 1341 academic and 5444 administrative staff composed of different ethnic backgrounds. The prevalence of components of metabolic syndrome including diabetes mellitus, hypertension and cardiovascular problems are increasing leading to high morbidity and mortality tolls among the staff since few years back [22,23].

A total of 704 study participants were randomly selected using their payroll as a sampling-frame and split randomly into training (equation, n = 352) sample and testing (validation, n = 352) sample. A sample size calculated for developing cut-off values for obesity in the same study participants was used for this analysis, the details methods is published elsewhere [28]. Employees of Jimma University were selected to serve as study population due to high ethnic diversity compared to the general community in the surrounding. Administrative and academic staff members who were actively working and were not away for more than one week during recruitment period were considered for inclusion into the study.

Participants who had physical disability including deformity (Kyphosis, Scoliosis), pregnant women, and limb deformity that prevents standing erect and those who were seriously ill during the study period were excluded.

3. Measurements

Data were collected using WHO STEPS Questionnaire [29] adapted to the local context. A stepwise approach to collect sociodemographic, anthropometric and body fat percent data was undertaken. The data were collected by five clinical nurses who were recruited based on the qualification needed for conducting data collection. A five days training was given to interviewers and supervisors on interviewing approach, anthropometric measurements and data recording before the actual data collection. All the measurements and interviews for both training and testing population were done under close supervision.

3.1. Anthropometry

Body-weight was measured to the nearest 0.1 kg with a digital scale of the Air Displacement Plethysmography (COSMED, Rome, Italy). The validity of the scale was checked using an object of a known weight every morning. Height was measured with an adjustable portable stadiometer which was accurate to 0.1 cm in a private place with the study participants wearing light clothing and their heads positioned at the Frankfurt Plane and the four points (heel, calf, buttocks and shoulders) touching the vertical stand and their shoes taken off. Before starting the measurements, the stadiometer was checked using a rod of known length. Body mass index (BMI) was calculated by dividing weight in kilograms by height in meters squared.

3.2. Body fat percent (BF %)

Body fat percent was determined using Air Displacement Plethysmography (COSMED, Rome, Italy) [30–32] after calibration of the machine for adults. The subjects wore a similar swimming pant and swimming cap covering the hair to prevent air trapping under clothing while all ADP tests were conducted.

At the beginning of each testing day, quality-control procedures were performed. The participants were asked not eat or exercise and drink coffee 4–5 h prior to the test. They were also told not to smoke or drink alcohol within 2 h, not to participate in vigorous/high intensity weight training 12 h prior the test and were given advice to come after resting. For each participant, age, sex, height and identifiers (ID) were entered into a computer. A two-step calibration procedure was then performed, first with the empty test chamber and then with a calibration cylinder. While the second calibration step was being performed, the subject was weighed on a calibrated electronic scale. Next, the participant was asked to sit inside the BOD POD chamber and instructed to remain still and continue normal breathing while the body volume was being measured. The measurement took two minutes and ADP was used in this study as the standard reference (±3%) of BF% [33].

3.3. Data analysis

The questionnaire was checked for completeness by the investigator every day. Data were edited and doubly entered into EpiData version 3.1 and then exported to cleaned and analyzed using SPSS for windows version 20. Descriptive analysis was used to describe the study subjects. Multivariable linear regression was fitted for the training population (n = 350) to develop BF% prediction equation for Ethiopian adults with ADP as dependent variable and BMI, age and sex as independent variables after checking all assumptions.

The BF% prediction equation developed on the training population was applied on the testing (validation) population (n = 352) for estimating BF%. Then, on the testing (validation) population (n = 352), BF% measured using ADP and the one estimated using the newly developed Ethiopian BF% prediction equation and the Caucasian BF% prediction equation were compared using paired t-test. All results were expressed as means with their standard deviations.

Validity measures (sensitivity, specificity, positive predictive values and negative predictive values), Kappa statistics and Bland Altman plot were used to determine the validity of the new equation and the Caucasian equation as compared with ADP measured BF%. Agreement between ADP measured BF% and the Caucasian prediction equation was also determined using Bland Altman Plot. For the Bland Altman plot, in both cases, the difference between ADP measured BF% and the one predicted suing the equations were plotted on Y-axis against the average on measured and predicted BF% on the X-axis. It is recommended that 95% of the data points should lie within the ± 2 SD of the mean difference and mean difference should be at zero point on the graph [34].

3.4. Ethical consideration

Before data collection, the study was approved by Institutional Review Board of Jimma University Collage of Health Sciences. Written informed consent was also obtained from the study participants after the purpose of the study was clearly explained to all study participants. The right of study participants to refuse participation or withdraw from the study at any point was respected. All data were kept confidential and the study participants were informed that the information they gave will not be disclosed to anyone. To assure complete confidentiality, other identifying information including name were not recorded on questionnaire.

4. Results

Comparison of the background characteristics of the training (equation) and testing (validation) sample showed that the two groups are comparable with regard to background characteristics. There was no a significant difference between the two groups in terms of sex, age, ethnicity, body mass index ADP measured body fat percent (Table 1). As presented in Table 2, a multivariable linear regression model showed that age, sex and BMI were significant predictors of ADP measured BF%. Accordingly, the BF% prediction equation of Ethiopian adults was generated as follows:

Table 1.

Background and anthropometric characteristics of the study training (Equation) and testing (validation) sample.

Variables	Study sample		P
	Training (Equation, n = 350)	Testing (Validation, n = 352)
	n (%)	n (%)
Sex
Male	152 (49.7)	154 (50.3)	0.932
Female	198 (50.0)	198 (50.0)
Ethnicity
Oromo	120 (47.4)	133 (52.6)	0.446
Amhara	102 (47.9)	111 (52.1)
Gurage	24 (63.2)	14 (36.8)
Kefa	25 (50.0)	25 (50.0)
Othersa	29 (60.4)	19 (39.6)
Dawero	28 (49.1)	29 (50.9)
Yem	22 (51.2)	21 (48.8)
Age, mean (±SD)	35.38 (±9.46)	36.05 (±9.28)	0.347
BMI, mean (±SD)	24.03 (±4.90)	24.21 (±4.62)	0.617
ADP measured Body Fat %, Mean (±SD)	31.83 (12.33)	32.38 (11.85)	0.547

SD: Standard deviation, ADP: Air displacement Plethysmography, BMI: body mass index.

^aSidama, Wolaita, Tigre.

Table 2.

Multivariable linear regression model predicting body fat percent of among the testing (equation) sample of Ethiopian adults.

Model	Unstandardized β Coefficients	Std. Err	P	95.0% CI for B
				Lower Bound	Upper Bound
(Constant)	−8.601	1.856	<0.0001	−12.251	−4.951
BMI	1.521	0.067	<0.0001	1.389	1.654
Age	0.243	0.033	<0.0001	0.179	0.307
Sex	−10.568	0.607	<0.0001	−11.762	−9.374

Ethiopian Body fat precept Equation: BF% = −8.601 + BMI (1.521) + Age (0.243) + Sex (−10.568).

Maximum VIF = 1.14, Adjusted R² = 0.79.

BF%: Body fat percent.

BF% = −8.601 + BMI (1.521) + Age (0.243) + Sex (−10.568), where sex = 1 for males and 0 for females.

Comparison of BF% measured by ADP and BF% estimated using the newly developed Ethiopian equation and Caucasian equation showed that there was no significant difference between ADP measured BF% and Ethiopian equation predicted BF% with a mean (±SD) difference of 0.03 (±5.44), P = 0.932. Conversely, there was a significant difference (<0.0001) between ADP measured BF% and the Caucasian Equation estimated BF% with a mean (±SD) of 6.83 (±5.57), Table 3.

Table 3.

Paired t-test comparing body fat percent measured by ADP and estimated using Ethiopian and Caucasian BF% prediction equation among the testing sample of Ethiopian adults.

Paired comparison		n	Mean	Std.	P
Pair 1	ADP measured BF%	352	32.38	11.85	0.932
	Ethiopian equation predicted BF%	352	32.35	10.52
	Difference	352	0.03	5.44
Pair 2	ADP measured BF%	352	32.38	11.85	<0.0001
	Caucasian equation predicted BF%	352	25.54	9.78
	Difference	352	6.83	5.57

BF%: body fat percept, ADP: Air displacement Plethysmography, Std: standard deviation.

Caucasian equation: BF% = ^1.294 * BMI ‡ 0.20* Age – 11.4* Sex – 8.0.

Ethiopian Equation: BF% = −8.601 + BMI (1.521) + Age (0.243) + Sex (−10.568).

The validity of Ethiopian and Caucasian BF% prediction equation in detecting obesity based on WHO cut-off for BF% (>25) was compared with ADP in the testing (validation) sample. It was observed that in males the Ethiopian equation performed very good to excellent with sensitivity (81.1%), specificity (91.9%), positive predictive value (90.9%) and negative predictive value (82.9%). Conversely, the Caucasian equation had poor sensitivity (17.5%) and negative predictive values (52.5%), while it demonstrated an excellent specificity (98.6%) and positive predictive value (93.3%). Likewise, there was a substantial Kappa agreement (K = 0.741) between Ethiopian equation and ADP in diagnosing obesity among males, while there was a slight Kappa agreement (K = 0.156) between Caucasian equation and ADP (P < 0.001).

Similarly, in females, the Ethiopian equation performed very good to excellent with sensitivity (87.4%), specificity (87.5%), positive predictive value (93.7%) and negative predictive value (76.7%). However, the Caucasian equation had poor sensitivity (47.4%) and negative predictive values (47.0%), while it demonstrated an excellent specificity (100.0%) and positive predictive value (100.0%). Likewise, there was a substantial Kappa agreement (K = 0.720) between Ethiopian equation and ADP in diagnosing obesity among females based on WHO cut-off for BF% >35, whereas, there was a fair Kappa agreement (K = 0.365) between Caucasian equation and ADP (P < 0.001), Table 4.

Table 4.

Validity of Ethiopian and Caucasian BF% prediction equation in detecting obesity among validation sample of Ethiopian adults as compared with the obesity determined using ADP.

Sex	Obesity based BF%	TP(a)	FP(b)	FN(c)	TN(d)	Total	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Kappa	Agreement	P
Males	Caucasian equation Predicted > 25	14	1	66	73	154	17.5	98.6	93.3	52.5	0.156	Slight	0.001
	Ethiopian equation Predicted > 25	60	6	14	68	154	81.1	91.9	90.9	82.9	0.741	Substantial	<0.001
Females	Caucasian equation Predicted > 35	64	0	71	63	198	47.4	100.0	100.0	47.0	0.365	Fair	<0.001
	Ethiopian equation Predicted > 35	118	8	17	56	198	87.4	87.5	93.7	76.7	0.720	Substantial	<0.001

Sensitivity = a/a + c, specificity = d/b + d, positive predictive value (PPV) = a/a + b, negative predictive value (NPV) = d/c + d.

Kappa agreement (0 = no/poor), (0.01–0.20 = slight), (0.21–0.40 = fair), (0.41–0.60 = moderate), (0.61–0.80 = substantial), and (0.81–1.00 = almost perfect) [35].

ADP measured body fat percentage >25 for males and >35 for females was used as a a confirmatory test [14].

A similar cut-off was used for body fat percent predicted using Ethiopian Equation and Caucasian prediction equation.

ADP: air displacement Plethysmography.

Bland Altman plot for the agreement between ADP measured BF % and the one predicted using Ethiopian equation showed a good agreement. As depicted in Fig. 1, the mean difference is close to zero (Mean = 0.03) and 95% of the data points fall within mean ± 2 SD.

Fig. 1.

Bland Altman plot showing the agreement between BF% measured by ADP and BF% estimated using Ethiopian prediction equation among the testing sample.

On the contrary, ADP measured BF% and the one predicted using Caucasian equation showed a poor agreement with a mean difference of 6.83. A shown in Fig. 2, the difference was greater than zero showing that the Caucasian equation underestimated BF% among Ethiopian adults and some data points are scattered outside the range of mean ± 2 SD.

Fig. 2.

Bland Altman plot showing the agreement between BF% measured by ADP and BF% estimated using Caucasian prediction equation among the testing sample.

Figure 3 shows a plot of BF% measured by ADP, and BF% Estimated using Ethiopian and Caucasian Equation in the testing sample. For all observation ranging from low BF% to high BF%, the new Ethiopian equation performed very close to the ADP measurement while the Caucasian equation underestimated BF% for all observations.

Fig. 3.

Relationship between ADP measured body fat percent with the Ethiopian and Caucasian equation predicted body fat percent among the validation sample of Ethiopian adults.

5. Discussion

This study developed body fat prediction equation on a training sample of Ethiopia adults based on BMI, age and sex, which demonstrated a very good to excellent performance on the testing sample in terms of validity measures (sensitivity, specificity and predictive values), a substantial Kappa agreement [35] and an agreement on Bland Altman Plot [34,36]. An ideal agreement in Bland Altman Plot is zero difference between measurements, which was observed in our data as the mean difference for the Ethiopian Equation is close to zero (mean difference = 0.03). There was no statistically significant differences (p > 0.05) between measured BF % and predicted BF% using Ethiopian prediction equation, showing the validity of the new equation in estimating BF%.

Although there have been numerous studies conducted to find out the best anthropometric indicator for detecting body fatness among different population groups [37–39], ethnic and racial differences in the predictive power of BMI based equation has been reported [5,38,40]. Accordingly, the need for developing ethnic specific prediction equations has been suggested to avoid such a problem [40]. This study also demonstrated that the Caucasian prediction equation significantly underestimates BF% among Ethiopian adults, which is consistent with the finding reported from a study among adult population of different ethnic groups [5]. It was also reported that body-mass index (BMI) may significantly misclassify subgroups with excess body fat among the different races and genders [3]. Different levels in energy intake, energy expenditure and body frame could be possible explanations for underestimation of BF % by the Caucasian equation [6]. Differences in body build have also been reported within Caucasian populations [5,37,38].

Our findings showed that the Caucasian prediction equation underestimated body fat percent of Ethiopian adults when validated against ADP, which is considered to be an accurate method in healthy adults [41 ]. The’ Bland Altman plot [34,36] also showed that the differences plotted on the Y-axis against the average of measured fat and predicted fat on the x-axis showed that 95% confidence interval did not include all the points and the mean difference is greater than zero showing the Caucasian prediction equation underestimates body fat percent.

Subjects with a small body frame, are likely to have a relatively lower fat free mass (due to lower muscle mass) compared to other people of same body height and hence BMI is likely to underestimate their body fat when the prediction equation developed in subjects with a bigger body build is applied [5,6,38]. The reason for the different relationships between body fat and BMI in the different populations is unknown. Apart from differences between dietary patterns and differences in physical activity, variations in body build may be an important contributor [38,39]. It has been shown that relation between body fat percent and MBI is curvilinear [42] indicating that BMI underestimates BF% even more for very obese subjects.

Differences in the relative leg length leads to differences in body fat percent for the same BMI such that people with lower sitting height to height ratio (people with longer legs relative trunk) have lower body fat percent [39]. It is known that there are differences in relative sitting height between Caucasians and Blacks, and between Caucasians and Asians, with blacks having relatively longer legs and Asians having, relatively shorter legs [37,40]. Apart from relative leg length, a stocky or slender body build may be one of the explanations for the difference [6]. A stocky person is expected to have more muscle mass compared to a slender person of the same body height [6,40]. As Ethiopians have slender body build, the Caucasian equation could underestimate their body fat percent. Thus, for the same BMI, the slender person will have more body fat [38,39]. Additional anthropometric measures may be necessary to improve the quality of the BMI as an indicator of body fatness among ethnic groups [5,38,39,43,44].

Given the aforementioned contexts, the findings of this study have wider practical implications. Assessing body fat percent at the community level is very critical in the wake of an increasing prevalence of obesity related to urbanization, dietary transition, increasing consumption of more processed foods and transition into motorized way of life. This needs a simple tool for early detection and public health interventions, as the facilities for determination of body fat percent like ADP are expensive and non-portable, which makes them unavailable for day to day service provision. Therefore, development of body fat percent prediction equation based on locally relevant data is a critical input to such prevention efforts. This is especially important as Ethiopia is aspiring to be a middle income country by 2025, life expectancy has risen to 64 years and there is an increasing prevalence of chronic non-communicable diseases related to obesity [45,46]. Prevention of NCDs will enable the country to reap its demographic dividend by avoiding untimely death and disability among working adults [47].

The current study has strength of using a validation sample to test the performance of the prediction equation in estimating body fat percent in a large sample. The study used local data for development of prediction equation and ADP as a confirmatory test for measuring body fat percent which are some of the strengths. Although getting sample of all ethnicities in Ethiopia was difficult given the non-portable nature of the ADP, the study tried to include the ethnicities that constitute the majority of the population.

6. Conclusion

The new Ethiopian BF% prediction equation performed very well in predicting BF% in the in terms of validity measures, Kappa agreement and Bland Altman plot; while the Caucasian equation significantly underestimated body fat percent among Ethiopian adults. The results imply that the new Ethiopian equation can be used as a cost effective and user friendly screening method for early detection of obesity and prevention of associated morbidity and mortality in Ethiopian adults.

Abbreviations:

BMI

Body mass index

BF%

Body fat percent

ADP

Air displacement Plethysmography

CVD

Cardiovascular disease

IRB

Institutional Review Board

WHO

World Health Organization

Availability of data and materials

All data used to support the findings of this study are available from the corresponding author upon request.