Estimation of CT-Derived Abdominal Visceral and Subcutaneous Adipose Tissue Depots from Anthropometry in Europeans, South Asians and African Caribbeans

Background South Asians and African Caribbeans experience more cardiometabolic disease than Europeans. Risk factors include visceral (VAT) and subcutaneous abdominal (SAT) adipose tissue, which vary with ethnicity and are difficult to quantify using anthropometry. Objective We developed and cross-validated ethnicity and gender-specific equations using anthropometrics to predict VAT and SAT. Design 669 Europeans, 514 South Asians and 227 African Caribbeans (70±7 years) underwent anthropometric measurement and abdominal CT scanning. South Asian and African Caribbean participants were first-generation migrants living in London. Prediction equations were derived for CT-measured VAT and SAT using stepwise regression, then cross-validated by comparing actual and predicted means. Results South Asians had more and African Caribbeans less VAT than Europeans. For basic VAT prediction equations (age and waist circumference), model fit was better in men (R2 range 0.59-0.71) than women (range 0.35-0.59). Expanded equations (+ weight, height, hip and thigh circumference) improved fit for South Asian and African Caribbean women (R2 0.35 to 0.55, and 0.43 to 0.56 respectively). For basic SAT equations, R2 was 0.69-0.77, and for expanded equations it was 0.72-0.86. Cross-validation showed differences between actual and estimated VAT of <7%, and SAT of <8% in all groups, apart from VAT in South Asian women which disagreed by 16%. Conclusion We provide ethnicity- and gender-specific VAT and SAT prediction equations, derived from a large tri-ethnic sample. Model fit was reasonable for SAT and VAT in men, while basic VAT models should be used cautiously in South Asian and African Caribbean women. These equations will aid studies of mechanisms of cardiometabolic disease in later life, where imaging data are not available.


Introduction
Escalating global levels of obesity will fuel an epidemic of diabetes and cardiovascular disease [1]. Deposition of central adipose tissue, and in particular the relation between visceral and subcutaneous fat, contribute to the development of cardiometabolic disease, yet their respective roles remain incompletely understood [2][3][4]. There are marked ethnic

Study sample
We used cross-sectional follow-up data from the Southall And Brent REvisited (SABRE) study, a community-based triethnic cohort study of men and women living in north-west London. The cohort's focus is on inter-ethnic differences in cardiovascular disease and diabetes, details have been published elsewhere [28]. All South Asian and African-Caribbean participants were first-generation migrants. Most African Caribbeans (92.5%) were born in the Caribbean and the remainder were born in West Africa. Most (82%) South Asians were born in the Indian subcontinent and 14% were born in East Africa, approximately half (52%) were of Punjabi Sikh origin. Ethnicity was confirmed by participants at interview. Participants aged 40-69 (n=4857) were randomly selected from age-and gender-stratified general practitioner lists and workplaces at baseline (1988)(1989)(1990)(1991), and were followed-up between 2008 and 2011, aged 58 to 85 years (n=4196). A total of 1410 people attended a research clinic at follow-up  and underwent CT scanning. Of these, 669 were white European (23% female), 514 were South Asian (15% female) and 227 were African-Caribbean (49% female), ( Figure  S1, supporting information). As the baseline study had been initially designed to study ethnic differences in cardiometabolic disease in men, a male preponderance in the data exists. This is absent in African Caribbeans, who were recruited later into the study, when the importance of cardiovascular disease in women was better recognised.

Anthropometric measurements
At the follow-up visit, measurements were undertaken by trained researchers in the study clinic at St Mary's Hospital, London. Height was measured using a stadiometer, with the participant barefoot and standing straight with the head level. Participants were weighed barefoot wearing a hospital gown, using a Tanita TBF-410 MA body composition analyser. This was used to calculate fat % and fat mass (in kg), in addition to total weight. Waist circumference was measured halfway between the costal margin and the iliac crest. Hip circumference was measured at the level of the greater trochanter. Thigh circumference was measured by instructing the participant to place their right foot on a chair, identifying and marking the midpoint between the hip crease and the patella, instructing the participant to return their leg to a straight position, and then measuring the circumference at the marked point. Circumferences were measured to the nearest millimetre using a fibre glass tape with a spring balance set to a constant tension of 600g. Body mass index (BMI) was calculated as weight in kg/ (height in m) 2 . We assessed the reliability of anthropometric measurements by comparing two sets of measurements from one observer on a sample of thirty participants. The coefficients of variation for weight, height and waist, hip and thigh circumferences were 0.16%, 0.23%, 1.98%, 0.71% and 7.50% respectively.

Radiological measurements
Visceral and subcutaneous adipose tissue (VAT and SAT) were measured by computer tomography (CT) scan at 125kV with a Philips MX 8000 IDT64 detector, which scanned a single slice of 10mm thickness at the fourth lumbar (L4) vertebral level. On average, the largest proportion of fat is seen at L4, most authors claim VAT at L4-5 level is best correlated with total VAT volume, and scans at this level best allow differentiation between VAT and SAT [29]. Images were taken with the participant supine, in fixed inspiration and with their arms extended overhead.
VAT was measured on the CT image by circumscribing the visceral compartment manually (using Image-J software [30]) and using an attenuation range of -190 to -30 Hounsfield units to quantify adipose tissue within. Total abdominal adipose tissue was measured using an automated function that gave the adipose tissue in the whole cross-sectional slice. SAT was calculated by subtracting the area of adipose tissue within outer border of the abdominal wall musculature from the total adipose tissue, using the above attenuation range (Figure 1).
The reliability of VAT and SAT measurements was tested by comparing two sets of measurements from one observer on a sample of thirty participants. The coefficient of variation for VAT was 0.95% and for SAT 0.68%.

Statistical analyses
Data were split randomly within gender and ethnic group, creating a derivation data sub-set (~ 66%), and a validation data sub-set (~ 33%). Summary statistics were generated by gender, ethnicity and derivation/ validation sub-set. ANOVA or chi-squared tests (for continuous/ categorical data respectively) were used to determine ethnic differences and differences between derivation and validation data sub-sets. Spearman's correlation coefficients were calculated for associations between VAT or SAT and age and anthropometrics in the derivation sub-set, by gender and ethnicity.
Two strategies were used to create prediction equations. Firstly, models were generated for VAT using age and waist circumference, and for SAT using age, weight and height. These models aimed to provide prediction tools when only basic anthropometry is available, as is likely to be the case in most research studies. Secondly, backwards stepwise regression was used with age (as a forced variable) and the following in the initial model; height, weight, waist, hip and thigh circumferences. Variables were eliminated sequentially until all variables in the model were significant at p<0.2. This process was carried out for each gender and ethnic group within the derivation dataset. We chose to represent weight, height, waist and hip circumferences as separate variables in models, rather than using body mass index (BMI) or waist:hip ratio. This was to better appreciate effects of each parameter on VAT/ SAT and to avoid multicollinearity in models. Age was specified a priori as a forced variable in all models as body fat distribution is thought to vary with age, with preferential VAT deposition with increasing age, particularly in women [31]. Goodness of fit was assessed with the adjusted R 2 statistic.
VAT and SAT prediction equations were internally crossvalidated, using the validation dataset to compare actual mean VAT or SAT values with mean VAT or SAT predicted by equations, and expressing these as proportions of the respective actual means. The difference was plotted for each group to enable visual inspection of model fit. Additionally, prediction equations from studies that used similar anthropometrics as predictors were applied to the validation dataset (for South Asian and African Caribbean participants separately, and then to the whole validation data subset) and differences between resulting actual and predicted mean VAT/ SAT values were compared [17,22,24,26,27,32] (table S1, supporting information).
Sensitivity analyses were performed (comparing R 2 values) on the basic models described for VAT and SAT (above), under the following conditions: the addition of total fat mass/ percentage values (from bioimpedance data)/ body mass index, participants with and without diagnosed diabetes at follow-up (defined as a diagnosis of diabetes/ prescription of anti-diabetic medications from primary care record review, or recall of physician-diagnosed diabetes plus either year of diagnosis or receipt of named anti-diabetic medication from participant questionnaire, or fasting/ oral glucose tolerance test plasma glucose results meeting WHO 1999 criteria at 2008-2011 follow-up [33]), and participants in the younger (56-72 years) and older (73-86 years) age groups.
All analyses were performed using Stata 12 (College Station, Texas).

Results
White European men were generally larger in all anthropometric and body composition parameters compared with South Asian men (table 1). However, after adjusting for age, height and total fat mass, South Asian men had the greatest, and African Caribbean men the smallest areas of visceral adipose tissue (VAT). South Asian men also had the greatest area of subcutaneous adipose tissue (SAT), followed by African Caribbean men, then white European men. African Caribbean women were the heaviest, with the largest waist, hip and thigh measurements, whereas South Asian women generally had the smallest measurements (table 2). Reflecting the trend in men, South Asian women had the greatest adjusted VAT and SAT areas, with African Caribbean women having the lowest VAT area, and white European women having the lowest SAT area. In both genders, there was evidence for inter-ethnic differences for most anthropometrics. Characteristics did not differ between the derivation and validation sub-sets in either gender (all p-values >0.05).
In univariate analysis, age and height were not consistently associated with VAT or SAT. Weight, BMI, waist:hip ratio and waist, hip and thigh circumferences were generally positively associated with VAT and SAT in both genders and all ethnic groups, with the strongest associations tending to be with waist circumference for VAT and weight or BMI for SAT. For VAT, anthropometric associations were mostly weakest in South Asians (table 3).

Prediction models
Waist circumference was used in VAT basic models, and height and weight in SAT basic models, as waist and BMI were the simple measurements most correlated with VAT or SAT respectively in most sex/ ethnic groups, table 3. We used height and weight in SAT models rather than BMI in order to better appreciate the relative contribution of each measure. Basic prediction models for SAT (age, weight and height: adjusted R 2 0.69-0.77) fitted better than those for VAT (age and waist: adjusted R 2 0.35-0.71). Model fit for VAT was better in men (adjusted R 2 range 0.59-0.71) than in women (adjusted R 2 range 0.35-0.59). Conversely, basic SAT models fitted better in women (adjusted R 2 range 0.75-0.77) than men (adjusted R 2 range 0.69-0.72) (table 4 for equations and table S2 of supporting information for full models).
Expanded models were constructed with age as a forced variable and height, weight, waist, hip and thigh circumferences as potential factors in a backwards stepwise regression. For VAT and SAT, expanded models improved adjusted R 2 when compared to basic models, especially in South Asian and African Caribbean women for VAT and African Caribbeans of both genders for SAT (table 4, table S2).
Basic and expanded models were cross-validated by comparison of mean actual and mean predicted values (for VAT and SAT) on the validation dataset, by gender and ethnicity. For VAT and SAT, differences between predicted and actual means were generally small, ranging from 1.2% to 9.3% of actual means for basic VAT models, 0.5% to 6.7% for expanded VAT models, 0.3% to 9.5% for basic SAT models and 0.2% to 8.1% for expanded SAT models. The exceptions to this were VAT models for South Asian women, which showed an 11.8% difference for the basic model and a 16.0% difference for the expanded model (see Figure 2). VAT prediction equations from previous studies were validated separately in our South Asian and African Caribbean populations, then in the whole validation data subset (comparing two equations derived from Indian populations living in India [26] [27], one from an African American population [22], and three from white European Canadian/ North American populations [17,24,32]. Only equations using similar anthropometrics (see table S1, supporting information for external equations in full) were compared to the expanded equations described in this study, using the validation dataset, see Figure 3. In South Asians, Brundavani et al's equations performed similarly to ours for men and women (i.e. difference in actual and predicted means as a proportion of the actual mean did not differ from that found in our cross-validation by >10%) [27]. Goel et al's equations did not perform as well for men (difference in means: Goel et al; 23.6% of actual mean, our equations; 5.6%) [26]. For African Caribbeans, the equations of Stanforth et al worked similarly to our equations in both genders [22]. Our equations performed better than those of Despres et al, Bonora et al and Ross et al in South Asian men and African Caribbeans of both sexes, though not in South Asian women [17,24,32]. When equations were applied across all participants in the validation dataset, our equations performed better than all externally-derived ones (difference in means: our equations; 0.6% of actual mean, externally-derived equations; 4.4% (Despres et al) to 22.8% (Goel et al)) [17,22,24,26,27,32].
Sensitivity analyses adding fat mass or BMI to models or deriving models in subgroups (people with diabetes or age < 73) did not generally show large differences in model adjusted R 2 values, and none that were replicated consistently across gender and ethnic groups (table S3 supporting information).

Discussion
In this large UK cross-sectional population-based study we provide novel and validated gender-and ethnic-specific equations for estimating abdominal visceral and subcutaneous fat deposition from anthropometric data. VAT has been consistently linked with conventional cardiovascular risk factors [34], emergent risk factors [35] and type 2 diabetes [36]. However, contention exists over whether SAT has a neutral, protective or detrimental role, as well as possible explanations for this [3,4]. Ethnic differences exist in fat deposition, with South Asian groups having more VAT and African Caribbeans more SAT, relative to white Europeans -these differences have been linked to differing cardiometabolic disease rates [6,37]. Therefore, inter-ethnic variation in patterns between fat deposition and cardiometabolic disease can help understanding of aetiology, identify mechanisms of disease and highlight pathways for intervention. Anthropometric measurements alone cannot distinguish between visceral and subcutaneous fat deposition, and are unreliable indicators of fat deposition in some ethnic groups [6,21]. Consequently, accurate quantification of VAT and SAT deposition is important to inform future research in this field.
We showed expected ethnic and gender differentials in CTquantified VAT and SAT i.e. greater VAT in men than women, greater SAT in women than men, greater VAT and SAT in South Asians than white Europeans and lower VAT and more SAT in African Caribbeans than white Europeans [5,22,38,39]. Univariate associations between fat depots and anthropometrics were generally weakest in South Asians, reflecting low correlations found in previous work [26].
Whilst variance explained by the prediction equations differed by fat depot, gender and ethnicity and use of simple or expanded models, goodness of fit for VAT prediction equations was similar to that found in previous studies, with authors reporting R 2 values of 0.51 [21] to 0.81 [32] for comparable models. Corresponding with the results from this study, previous work has found that VAT models' goodness of fit was  [26], and in women of all ethnic groups studied [18], [22], [27]. Other studies have reported high R 2 values for SAT prediction models (0.67-0.85) [16,18,25,26], in keeping with this study, and where reported, these have been higher for women. Cross-validation of our models, by comparison of actual and predicted means, suggested VAT models performed well in all groups except for South Asian women. However, for basic equations, the difference in the means was 21.4 cm 2 , or 11.8% of the actual mean, which compares favourably with results from equations derived by Goel et al, who reported a difference in means of 18.6 cm 2 (15.8% of the actual mean), and Bonora et al, who reported a difference in means of 36.0 cm 2 (21.3% of the actual mean) [17,26]. Differences between the true and estimated mean values of both depots were not numerically large, and were similar or an improvement on those reported by other studies. Where possible, we applied other authors' VAT prediction equations to our validation dataset. Equations derived from South Asian populations and the equation derived from an African American population performed similarly to our equations in these ethnic groups, with the exception of Goel et al's equations, which performed less well in South Asian menpossibly due to smaller numbers (n=95 for men) and a younger age range (18-50 years) of participants used to derive their equations [22,26,27]. As expected given their derivation from the same sample, our equations performed markedly better than any of the other equations tested when applied to the entire validation dataset [17,22,24,26,27,32]. Nevertheless, the comparison demonstrates the importance of gender-and ethnicity-specific equations in improving accuracy of VAT prediction at a population level.
This study has a number of strengths. To our knowledge, ours is the largest study to derive VAT prediction equations (n=1410), with only 2 other studies exceeding 200 participants (Brambilla et al: n=407, Stanforth et al: n=692) [20,22]. This is the first study to provide group-specific equations, rather than to include ethnicity as an adjustment factor in equations, which enables between ethnic group comparisons. We derived equations from easily and commonly measured anthropometrics; many other studies have presented equations derived from sagittal diameters or skinfold thickness [21,24,25] which may not be available in the majority of available datasets. Advantages of the methods employed include use of "generous" p-values for inclusion into complex models (P<0.2), to make use of effects which may not have high levels of statistical significance but may contribute importantly to VAT or SAT variation. We cross-validated our equations, and those of other studies, using a validation sub-set of the sample. Several previous studies have internally cross-validated their equations [17,19,20,22], but as far as we are aware, only 2 have compared them to equations derived from other studies [25,32] -both of which contained relatively small numbers of white male participants only. Although the sample was population-based, caution is needed when generalising results to different ethnic groups, i.e. most South Asians in this study were of Indian origin, therefore results may not be generalisable to other groups originating from the Indian subcontinent, e.g. Bangladeshis or Sri Lankans. For the basic VAT models, goodness of fit was poor for South Asian and African Caribbean women (R 2 0.35 and 0.43 respectively). Therefore it is recommended that the more expanded equations be used, if possible, in these groups. This study pertains to individuals aged only 56 and above, therefore the pre-menopausal hormonal influence on adipose tissue deposition in women is not accounted for in prediction equations [40]. However, since cardiometabolic disease is more common in mid to later life, and many population studies focus on this age group, we do not consider this to be a major drawback of this work. Maislin et al have recently suggested VAT quantification at L2-3, rather than L4-5, more accurately reflects total VAT volume [41]. However their findings were from an obese white European population and are unlikely to apply to our population. Errors in anthropometric measurements may have affected the reliability of our prediction equations. However, as the coefficients of variation for repeated anthropometric measurements on a sub-sample were low, we do not believe measurement inaccuracies will have substantially affected the equations' reliability.
In summary, this is the first UK-based population study to present ethnicity and gender-specific VAT and SAT prediction equations using routine anthropometrics in Europeans, South Asians and African Caribbeans. Model fit was reasonable for all ethnic groups for SAT and for VAT in men, while basic VAT models should be used cautiously in South Asian and African Caribbean women. Where imaging data are not available, these equations can aid studies of mechanisms of cardiometabolic disease in all ethnic groups.

Figure 3. Inter-study differences between actual and predicted mean VAT in: a) South Asian men, b) South Asian women, c) African Caribbean men, d) African Caribbean women and e) men and women of all ethnicities.
Inter-study VAT crossvalidation was performed in the validation dataset. Results from this study show predicted means from equations using age, height, weight, waist / hip/ thigh circumferences as predictors. For Figure 3a)