Visceral fat area is the measure of obesity best associated with mobility disability in community dwelling oldest-old Chinese adults

We used data from the SG90 study, which is nested within the population-based Singapore Chinese Health Study [17] (Supplementary Figure S1, Additional file 1). The latter is an on-going prospective cohort study designed to evaluate the genetic, dietary, and environmental determinants of chronic diseases in Chinese adults living in Singapore. Detailed descriptions of the study have been reported previously [17]. In brief, 63,257 participants (27,959 men and 35,298 women) aged 45–74 (mean 53) years old were enrolled between April 1993 and December 1998. Study participants belonged to either of the two major dialect groups of Chinese in Singapore (Hokkiens or Cantonese), and they were permanent residents or citizens who resided in government-built housing estates – where 86% of all Singaporeans resided during that period. After recruitment, the participants were re-contacted for follow-up interviews in 1999–2004 (phone), 2006–2010 (phone), and 2014–2017 (face-to-face) (Supplementary Figure S1, Additional file 1). The study was approved by the Institutional Review Board at the National University of Singapore, and written informed consent was obtained from all study participants.

There were 1305 participants who had hitherto participated in all three of the follow-up interviews, and who were also 85 years and older during the recruitment period of this study from July 2017 to August 2018. In this SG90 sub-cohort, we recruited the first 1000 consenting participants for in-person interviews that examined specific measures relevant to ageing, such as cognitive performance, quality of life, functional independence, etc. (Supplementary Figure S1, Additional file 1). During these interviews, we also asked participants for their sociodemographic characteristics and history of physician-diagnosed diseases that included diabetes, stroke, coronary artery disease, arthritis, and osteoporosis/hip fracture. We used the Singapore-modified Mini-Mental State Examination (SM-MMSE) [18] to screen participants for cognitive impairment. We also measured each participant’s handgrip strength and usual gait speed in line with methods described by the 2019 consensus update from the Asian Working Group for Sarcopenia (AWGS 2019), and then used the recommended cut-off values to detect cases of sarcopenia [19].

Five hundred and eighty-nine of these participants additionally consented to undergo measurements for height, weight, and waist circumference, as well as to an assessment of body composition using bioelectrical impedance analysis (BIA). Each participant was visited by a trained interviewer and had their height measured to the nearest 1 cm using a stiff, self-retracting, metallic tape measure, and their weight measured to the nearest 0.1 kg with a Soehnle Exacta Comfort digital weighing scale (Model S63315 PSD). BMI (kg/m²) was computed as weight divided by height squared. Waist circumference was measured by positioning a soft, flexible, plastic tape measure at a level that was 2.5 cm (1 in.) above the participant’s navel [20].

We used the InBody S10 body composition analyser (Biospace Co. Ltd., Seoul, South Korea) to assess body composition through direct segmental multi-frequency bioelectrical impedance analysis (DSM-BIA). Each subject was asked to maintain a supine posture for at least 10 to 15 min before measurements were commenced. Eight tetrapolar adhesive electrodes (two for each hand and foot) were then attached to the participant, and they were used to obtain 30 impedance measurements across six frequencies (1, 5, 50, 250, 500, and 1000 kHz) as well as 15 reactance and phase angle measurements across three frequencies (5, 50, and 250 kHz) at each of five body segments (left arm, right arm, trunk, left leg, and right leg). These resulting measurements were subsequently used to estimate appendicular skeletal muscle mass [19], percent body fat, and visceral fat area for each participant.

During the in-person interviews, participants were assessed for mobility disability using the Loco-Check questionnaire. This questionnaire was originally designed by the Japanese Orthopaedic Association as a self-reported checklist comprising seven physical activities related to daily living to screen for locomotive syndrome – defined as a condition of reduced mobility due to impairment of the locomotive organs [21]. Subsequent studies have shown that the Loco-Check questionnaire is also useful in estimating the extent of overall physical dysfunction in older adults [22, 23]. Participants were scored according to the number of activities that they reported having difficulty with, and those with a total score of four or more were considered to have mobility disability [24].

We first evaluated which of the four measures of obesity had the best performance in discriminating between oldest-old adults with and without mobility disability through in-sample receiver operating characteristic (ROC) analysis. In this analysis, we first constructed a classification model for each of the four obesity measures using a logistic regression model which included sex and the respective obesity measure on a continuous linear scale as the exposures of interest, and mobility disability as the outcome of interest. This classification model was trained on the entire dataset, and its output (the estimated probability of mobility disability) was then tested against the observed outcome of mobility disability for every participant in the dataset. Using the ‘roctab’ and ‘roccomp’ commands in Stata [25], in-sample ROC curves were created by plotting the true positive rate (sensitivity) against the false positive rate (1 − specificity) as the discrimination threshold of the classification model was varied. Finally, we compared the areas under the in-sample ROC curves (in-sample AUC_ROC) for waist circumference, percent body fat, and visceral fat area against that of BMI as a reference using an algorithm by DeLong et al. [26] through the ‘rocgold’ command in Stata [25], and further adjusted for multiple comparisons using Bonferroni correction. To further evaluate how these models might generalize to new cases, and to obtain more realistic estimates of their predictive performance, we also performed five-fold cross-validation using the ‘cvauroc’ package [27] in Stata [25]. In this cross-validation analysis, the dataset was first randomly partitioned into five equal-sized groups (folds). Then, for each of the four obesity measures, a classification model was trained on four of the five folds, and its output was then tested against the observed outcome for every participant in the remaining fold, creating a cross-validated ROC curve. This process was repeated five times for each obesity measure, until each of the folds had been used as the held-out test set, and five different cross-validated ROC curves had been generated. The mean area under these five cross-validated ROC curves was then reported for each obesity measure. As a sensitivity analysis to account for possible non-linear relationships between the obesity measures and mobility disability, we repeated both the in-sample and cross-validated ROC analyses, but modelled each obesity measure with a restricted cubic spline with three knots at its 10th, 50th, and 90th percentiles [28] using the ‘mkspline’ command in Stata [25]. We also evaluated the statistical significance of the non-linear spline term in each logistic regression model to assess for the presence of a non-linear association. For each obesity measure, we used the ‘roccomp’ command to compare the area under the in-sample ROC curve of the continuous linear model against that of its respective spline-transformed counterpart.

After visceral fat area had been identified as the measure of obesity with the highest discriminative performance, optimal cut-off points for men and women were derived separately by selecting the points with the highest Youden’s J statistic [29, 30] that also had at least 50% sensitivity and 50% specificity in sex-specific ROC curves. These cut-off points were used to define obesity by visceral fat area in a binary fashion. For additional comparisons, previously established cut-off points were also used to define cases of obesity through BMI (cut-off point of 27.5 kg/m²) [31], waist circumference (cut-off point of 90 cm for men and 80 cm for women) [32], and percent body fat (cut-off point of 30% for men and 40% for women) [33, 34]. Multivariable logistic regression models were used to compute odds ratios (OR) and 95% confidence intervals (CI) for the association between these four definitions of obesity and mobility disability. In Model 1, we adjusted for age, sex, and level of education (no formal education, primary, secondary and above). In Model 2, we additionally adjusted for other chronic conditions that could have contributed to mobility disability in older adults, such as diabetes, stroke, coronary artery disease, arthritis, osteoporosis/hip fracture, cognitive impairment, and sarcopenia. In Model 3, we further adjusted for the three respective measures (BMI, waist circumference, percent body fat, or visceral fat area) that were not utilized in that particular definition of obesity as continuous variables [35]. All statistical analyses were conducted using Stata/SE 14.2 software [25] and R version 3.5.3 software [36]. All P-values presented were two-sided, and P < 0.05 was considered statistically significant.