The contribution of musculoskeletal factors to physical frailty: a cross-sectional study

Participants completed questionnaires that documented physical activity [17] dietary information [18], lifestyle factors and health behaviours, sociodemographic factors, medications, comorbidities and post-baseline fractures that were radiologically confirmed.

Frailty was identified using a modified Fried frailty phenotype [6], which categorised individuals into frail, pre-frail or robust groups, based on the responses to five criteria including unintentional weight loss, weakness, low physical activity, exhaustion and slowness. Weakness was determined using handgrip strength (HGS) measured with a hand-held Jamar dynamometer (Sammons Preston, Bolingbrook, IL, UK) for women [19] and Vernier dynamometer (Venier Software and Technology, Beaverton, USA) for men. Vernier values were converted to Jamar equivalents using an equation previously described [20]. HGS values below cut-points equivalent to the lowest 20% stratified by sex and body mass index (BMI) were considered as weakness [6]. The HGS procedure was demonstrated to participants before the trials. For women, there was no interval between trials, and for men, there was a 3 s interval between the trials; for both sexes the mean of the maximum reading from each hand was used in analyses. Slowness was measured using the Timed Up & Go (TUG) test that measures functional mobility including balance and muscle performance [21]. The time was recorded for the participant to stand from a chair (standard height without armrests), walk 3 m, turn around and return to sit back in the chair. A score ≥ 10s was considered as slow [22]. Unintentional weight loss, exhaustion and low physical activity were self-reported. Low physical activity was ascertained using a single question that asked about the current mobility status of participations. Low physical activity was indicated when participants selected one of the following responses: “limited”, “inactive”, “chair” or “bedridden”, or “bedfast”. Participants with at least three of the five items in the modified Fried tool were categorised as frail, 1–2 items as pre-frail, and zero items as robust.

Using electronic scales and a wall-mounted Harpenden stadiometer, weight and height were measured to the nearest 0.1 kg and 0.1 cm, respectively, and BMI was calculated (kg/m²). Using dual-energy X-ray absorptiometry (DXA, Lunar Prodigy Pro), areal bone mineral density (BMD) was measured at the femoral neck, and measures of total and regional body fat and lean tissue mass were derived from the whole-body scans. DXA-derived lean mass was used as a surrogate measure for skeletal muscle mass. Appendicular lean mass index (ALMI) was calculated as appendicular lean mass divided by height squared (kg/m²); similarly, fat mass index (FMI) was calculated as whole body fat mass divided by height squared (kg/m²). Quantitative calcaneal ultrasound was performed on the left heel using a Lunar Achilles Insight ultrasonometer (GE Lunar, Madison, WI, USA), with measures expressed as speed of sound (SOS), broadband ultrasound attenuation (BUA) and stiffness index (SI) indicating bone quality [23]. Low trauma fractures (excluding fingers, toes, skull and face) from baseline visit until the follow-up visit for participants were ascertained using radiological reports from the radiological imaging centres servicing the region. This method of fracture ascertainment has been previously validated and described in detail [24, 25].

Comorbid conditions were classified using a modified Charlson comorbidity index (CCI) [26], that took into account the number and seriousness of comorbid diseases. Using a combination of self-reported and measured data, participants were assigned weights for the following conditions: myocardial infarction, congestive heart failure, peripheral vascular disease, cerebrovascular disease, dementia, chronic pulmonary disease, connective tissue disease, ulcer disease, moderate or severe liver disease, mild liver disease, diabetes hemiplegia, moderate or severe renal disease, diabetes with end organ damage, any tumour, leukaemia, lymphoma, metastatic solid tumour, and acquired immunodeficiency syndrome (AIDS) [26].

Socioeconomic Indexes For Areas (SEIFA) values using 2011 Australian Bureau of Statistics (ABS) census data were determined using ABS software (these rank areas in Australia according to relative socio-economic advantage and disadvantage), which were ranked in quintiles [16] and, due to small numbers, collapsed into three groups of low (quintile 1 and 2), medium (quintile 3) and high (quintiles 4 and 5). Data regarding current smoking status (yes/no) were derived from a questionnaire. Alcohol consumption exceeding 30 g/day was derived from the Victorian Cancer Council Food Frequency Questionnaire [18].

Descriptive characteristics of participants were presented as median (IQR) or mean (±SD) or n (%). Participants’ characteristics for frail, pre-frail, and robust groups were compared using analysis of variance (ANOVA) for parametric data, Kruskal-Wallis test for non-parametric data and chi-square test for categorical factors.

Exploratory analyses that investigated the contribution of muscle and bone parameters to frailty included binary logistic regression models where dichotomised frailty (yes/no; frailty vs pre-frail and robust combined) was considered the outcome. The area under the receiver characteristic curves (AUROCs) was also calculated. In addition, ordinal logistic regression models were used, where frailty was considered in three groups (frail, pre-frail and robust). For these statistical models, BMD, BUA, SOS, SI, ALMI and HGS measures were converted to standard deviation (SD) units using published reference data [19, 27‐30]. The following characteristics were included in multivariable models: age, weight and height or BMI, smoking status, alcohol, prior fracture, FMI, SES and CCI and retained if p < 0.05. Somers’ D nonparametric ordinal association was calculated as a post-hoc measure in ordinal logistic models. These confounders were included as studies have shown association between these factors and frailty [6, 11, 31‐33]. The estimated Somers’ D associations were adjusted for the characteristics included in the model [34]. Values closer to ±1 for Somers’ D suggest the higher contribution of a musculoskeletal factor in relation to frailty status and values tending towards zero in either direction indicate lower contribution of a musculoskeletal factor. Non-linearity assumption of the logistics models was assessed for key continuous covariates through comparing the linear trend logistic model with alternative nonlinear model assuming square and cubic nonlinear structure [35]. Boxplots were also investigated to explore the association between the frailty items and musculoskeletal parameters. Data for men and women were analysed separately. Minitab (Version 18, State College, PA, USA) and STATA (Version 16, College Station, Texas, USA) statistical software were used for statistical analyses.

The participants’ descriptive characteristics for the whole group and according to frailty status are displayed in Table 1. Frail men were older, had a higher mean BMI and were more likely to have higher mean FMI. Frail men were also under-represented in the highest SES tertile. No other differences were observed.

In unadjusted binary logistic models, SI, ALMI and HGS were associated with frailty in men (Table 2). The association was sustained after adjustments for relevant characteristics, where men with higher measures were less likely to be frail; SI (OR 0.73 95%CI 0.53–1.01, p = 0.06), ALMI (OR 0.48 95%CI 0.34–0.68, p < 0.001) and HGS (OR 0.11 95%CI 0.06–0.22, p < 0.001) (Fig. 2). No other associations were observed. Unadjusted ordinal logistic models showed similar results and these associations were sustained in adjusted models (Table 4).

Muscle parameters contributed more to frailty than bone parameters (Tables 2 and 4). HGS made a greater contribution to frailty than ALMI (AUROC 0.85 and 0.77 respectively, p < 0.01) (Table 2). This was also observed in the ordinal regression model, where HGS had a stronger effect (Somers’ D correlation of − 0.49 (− 0.57- (−) 0.41), p < 0.01) (Table 4), meaning that HGS improved our prediction of frailty status by 49%, greater than the 36% observed for ALMI. Of the bone parameters, only SI contributed to the model and the contribution after adjusting for age and BMI was 0.71 (AUROC) (p = 0.06).

For femoral neck BMD, lower mean scores were associated with the following frailty items; low physical activity, fatigue, slowness and low HGS. Weight loss was associated with higher mean femoral neck BMD. Similar results were observed for ALMI and heel ultrasound measures (BUA, SOS, SI), where lower mean scores were associated with low physical activity, fatigue, slowness and low HGS, while the inverse was true for weight loss. (Supplementary Fig. 1). Overall, for low physical activity, SI median scores had the greatest percentage decrease (13.0%) in median score between the yes and no groups compared to the other musculoskeletal parameters. For fatigue, ALMI had the greatest percentage decrease (1.7%) while for slowness, SI had the greatest percentage decrease (10.1%). For low HGS, ALMI had the greatest percentage decrease (5.7%). Lastly, for weight loss, SI had the greatest percentage increase of 7.4% between the no and yes groups. In correlation tests for frailty items, slowness (as measured by TUG) showed a strong correlation (Supplementary Table 1) and thus, TUG was modelled as binary logistic regression models. The TUG models were inferior in the AUROC for the ALMI model (from 0.77 to 0.76 (p < 0.001)) and the HGS model (from 0.85 to 0.79 (p < 0.001)).

Frail women were also older and shorter and had a higher mean BMI than pre-frail and robust women. They were more likely to have a prior fracture and a higher number of comorbid conditions (Table 1).

In unadjusted binary logistic models, BMD and HGS were associated with frailty (Table 3). After adjustments for age, weight and height, only HGS was associated with frailty, where women with greater HGS were less likely to be frail (OR 0.30, (95%CI 0.20–0.45) p < 0.001) (Fig. 2). No other associations were observed (Table 3). In unadjusted ordinal logistic models, HGS was associated with frailty, while an association that did not achieve statistical significance was observed with ALMI. After adjustments for age, height, FMI and SES, HGS and ALMI were associated with frailty (Table 4).

In binary logistic regression models, only HGS was observed to contribute to the model (AUROC 0.83, p < 0.001) (Table 3). However, when frailty was considered as three groups, ALMI was also observed to contribute to the model (p < 0.001). In these ordinal regression models, HGS had a stronger contribution to frailty than did ALMI (Table 4).

Overall boxplots that explored the association between the frailty items and musculoskeletal parameters did not reveal any apparent associations or trends across the musculoskeletal parameters. (Supplementary Fig. 2).

In correlation tests for frailty items, slowness (as measured by TUG) showed a strong correlation (Supplementary Table 2) and thus, TUG was modelled as binary logistic regression models. The TUG models were inferior compared with the HGS models (AUROC decrease from 0.83 to 0.74 (p = 0.001)).

For all models with significant exposure (i.e., for men: SI, ALMI and HGS and for women: ALMI and HGS), an inclusive model that included important potential confounders irrespective the p-values were employed. A link-test was utilised to test for any potential regression model specification error [36]. The link-test tests the idea that if a regression equation is properly specified no additional independent variables that are significant can be found except by chance. In addition, the square and cubic terms were added to the logistics models to test for potential nonlinearity. In conclusion none of the link-tests or testing the joint square and cubic terms were significant, and as such we did not detect any strong evidence of lack of regression goodness of fit. Furthermore, multivariate regression models that included necessary covariates were constructed, however, the conclusion did not substantially change by these adding additional factors (Supplementary Table 2-14).