Back Extensor Strength as a Potential Marker of Frailty Using Propensity Score Matching and Machine Learning

1. Introduction

Frailty is a geriatric condition characterized by increased vulnerability to external stressors [1]. It has been associated with adverse health outcomes, including higher mortality rates, increased risk of falls, and chronic diseases [2]. In 2001, Fried proposed a frailty assessment using five physical components, including grip strength [3]. Subsequently, various frailty classifications have been developed and implemented in clinical and research settings [4,5]. However, Fried’s frailty remains widely used and recognized for frailty assessment [6,7].

After Fried proposed grip strength as a frailty criterion, it has been shown to be an objective and reliable predictor of health outcomes [8,9]. However, it is important to note that grip strength is influenced by various factors beyond aging-related changes. Lifestyle factors, such as nutrition and depression status, have also been shown to impact grip strength [10,11]. Studies conducted in eleven European countries have revealed that even economic crises, such as a decrease in gross domestic product (GDP), can be associated with reduced grip strength [12]. These findings highlight the multifactorial nature of grip strength and emphasize the need to consider a comprehensive range of determinants when interpreting its significance as a frailty marker. Grip strength is not solely reflective of age-related decline but is influenced by a complex interplay of individual, societal, and economic factors.

As an alternative approach, researchers have studied muscle strength in different body parts [13]. By investigating muscle strength beyond hand grip, they aim to capture a more comprehensive understanding of an individual’s physical function and overall frailty status. Furthermore, the development of newly advanced dynamometers has enabled a wide range of muscle strength measurements, providing new possibilities for assessing muscle strength in multiple parts of the body, not just limited to hand grip strength [14,15].

The back extensor muscles play a crucial role in maintaining posture, stability, and overall trunk function [16]. Weakness or dysfunction in these muscles may contribute to balance issues, gait abnormalities, and increased vulnerability to falls, which are key components of frailty [17]. Understanding the impact of back extensor strength on frailty could have significant clinical implications.

The aim of this cross-sectional study was to explore the relationship between back extensor strength and trunk muscle/fat compositions, with a specific focus on investigating the potential of back extensor strength as an alternative marker of frailty. This study utilized a specifically designed chair equipped with a portable dynamometer to assess the strength of the back extensors. We anticipated a correlation between frailty status and back extensor strength, which could be explained by trunk muscle/fat composition. By investigating these associations, the study aimed to contribute to the understanding of the role of back extensor strength in frailty and its potential utility as a frailty marker.

2. Materials and Methods

2.1. Participants

For participant enrollment, a total of 446 eligible participants were selected from the Farmers’ Cohort for Agricultural Work-Related Musculoskeletal Disorders study conducted from 2013 to 2014, and 339 participants were selected from individuals involved in farming. The characteristics of participants have been detailed in previous study [18]. In consideration of the physically demanding nature and varied ergonomic tasks involved in farming [19,20] that contribute to specific changes in muscle volume, individuals who were not engaged in farming were excluded from the study. As a result, 590 individuals actively involved in farming participated in this study [21]. To mitigate the influence of external factors that could cause a decline in muscle volume, we excluded 21 participants with a history of spine surgery, as it is known to induce para-spinal muscle atrophy [22]. Additionally, nine participants with missing back extensor strength data were not included in the cross-sectional analysis, resulting in a total of 560 participants in the final analysis (Figure 1).

2.2. Measurements

2.2.1. Frailty Score

Frailty was measured based on Fried’s five criteria. Each criterion had values of 1 or 0: weight loss (weight loss in the past year ≥ 4.5 kg), self-reported exhaustion (at least three days a week), low physical activity (metabolic equivalent within the lowest 20%, adjusted for sex), slowness (usual 4-m gait speed within the lowest 20%, adjusted for sex and height), and weakness (grip strength within the lowest 20%, adjusted for body mass index and sex). Based on the number of satisfied criteria, participants were divided into two groups, non-frail (frailty score of 0–2) and frail (frailty score of 3–5).

Exhaustion was defined by answering the following question from the Center for Epidemiological Studies Depression Scale: “I felt all I did was an effort” for three or more days per week. Physical activity was assessed using the Korean version of the International Physical Activity Questionnaire [23]. To determine slowness, participants were divided into four groups based on sex and the median value of height, and the lowest 20% of the 4-m gait speed was calculated for each group and used as the cutoff value for slowness. Grip strength was measured three times on both hands, and the mean of these three values was calculated for each hand. The higher mean grip strength from both hands was selected for further analysis. To determine the weakness of grip strength, participants were divided into four groups based on sex and the median value of body mass index (BMI), and the lowest 20% of the grip strength was calculated for each group and used as the cutoff value for weakness.

2.2.2. Back Extensor Strength

The isometric strength of the back extensor was measured using the PrimusRS system (BTE Technologies Inc., Hanover, MD, USA) while participants were seated in a high chair to prevent their feet from touching the floor [24]. The height of the anchoring cable was modified to the height of the participants’ T7 spinous process level, and a seatbelt was used to stabilize the participants’ thighs. Participants were asked to push back as far as possible, which took 3 seconds with isometric back extension. The cable anchored to the chest harness was pulled, and the isometric strength was measured. This process was repeated five times, and the mean of the three median values (excluding the maximum and minimum values) was used.

2.2.3. Trunk Muscle/Fat Mass

The trunk muscle/fat mass measurement protocol in this study has been previously explained [25]. Ten consecutive computed tomography (CT) images at the mid-L4 vertebral level were acquired using a Philips MX 8000 IDT CT scanner (Philips Medical Systems, Cleveland, OH, USA), with exposure at 200 mAs, tube voltage set at 120 kV, and 1 mm thickness of slice. Total muscle mass (TMM, cm³) and total fat mass (TFM, cm³) in the trunk region were obtained using image processing software (Extended Brilliance Workspace version 4.5.3; Philips Healthcare, Best, The Netherlands). Pre-defined radiation attenuation ranges were used to demarcate adipose (from -190 HU to -30 HU) and muscle (-29 HU to +150 HU) tissues. TMM was subdivided into back muscle mass (BMM, cm³), psoas muscle mass (PMM, cm³), and abdominal muscle mass (AMM, cm³). BMM and PMM were derived manually from the TMM. AMM was obtained by subtracting PMM and BMM from TMM. The BMM consisted of the multifidus, iliocostalis lumborum, longissimus, and quadratus lumborum muscles. Visceral fat mass (VFM, cm³) was calculated by manually outlining the inner abdominal wall, and subcutaneous fat mass (SFM, cm³) was calculated by subtracting the VFM from the TFM. To reduce bias, one technician performed all scan and image processing procedures.

2.3. Statistical Analyses

All analyses were performed using R version 4.2.2. (The R Foundation, Vienna, Austria), Windows 10 [26]. Statistical significance was set at p<.05. The baseline characteristics of the study participants were presented and compared in two ways: between men and women; and between age < 65 years and age ≥ 65 years. Categorical variables are expressed as numbers and percentages and were compared using the chi-squared test. Continuous variables were expressed as the mean and standard deviation and were compared using Student’s t-test. A linear regression analysis was performed to examine the relationship between trunk muscle/fat composition and back extensor strength. We analyzed the presence of multicollinearity using the variance inflation factor (VIF), with covariates having absolute values of VIF > 10 indicating multicollinearity [27].

To determine the relative importance of covariates in linear regression, relative weight analysis (RWA) was calculated [28]. The absolute value of the RWA was used to indicate the correlation between the covariate and back extensor strength, with positive and negative RWA values indicating a positive and negative relationship with back extensor strength, respectively.

2.3.1. Propensity Score Matching

The participants were divided into two groups based on their back extensor strength, with the lowest 20% strength considered as one group and the rest as the higher group. Since the result of descriptive analysis, as shown in Table 1 and Table 2, indicated a significant relationship between sex, age, frailty, and back extensor strength, we performed propensity score (PS) matching of age and sex. The purpose of PS matching was to reduce the confounding effects by adjusting for confounding factors. PS is the conditional probability of an individual being in a certain group, based on covariates or predictor variables in the model [29]. Many researchers have used PS matching in medical and social science studies to reduce selection bias [30,31]. The nearest neighbor (NN) with a PS caliper, which imposed a tolerance level on the maximum PS distance, was used to match the two groups to reduce the risk of bad matches when the closest neighbor is located far away [32]. The matching ratio between the two groups was set as high as possible to increase the number of participants in the analysis. The standardized mean difference (SMD) of covariates was used to evaluate how well the two groups were paired, and an SMD less than 0.1 was accepted as the confounding bias was significantly reduced [33].

2.3.2. Multivariate Logistic Regression

To determine the potential of back extensor strength as a frailty parameter, we conducted a multivariate logistic regression analysis of frailty. We included age, sex, and BMI as covariates in frailty prediction, as these factors are commonly used covariates in strength assessments of frailty [34]. Odds ratios (OR) and 95% confidence intervals (CI) were calculated.

2.3.3. Extreme Gradient Boosting

To demonstrate the possibility of using back extensor strength as a frailty parameter, we performed Extreme Gradient Boosting (XGBoost) [35]. XGBoost is a novel gradient-boosting algorithm that has been proven to enhance the computing power of classification and regression models. The XGBoost model for frailty prediction was performed in two steps: (1) the grid search method was used to determine the optimal XGBoost hyperparameters, and (2) the optimal XGBoost model was repeated 100 times to calculate the CI of the performance parameters.

First, we used the grid search method to optimize the XGBoost hyperparameters (maximum depth of a tree, learning rate, gamma, subsample ratio of the training instance, minimum sum of instance weight, and subsample ratio of columns). For the grid search, the maximum depth of a tree was set to 2, 4, 6, 8, 10, and 12; the learning rate was set to 0.001, 0.0025, 0.005, 0.01, 0.02, 0.04, 0.06, 0.08, 0.1, 0.2, and 0.5; gamma was set to 0, 0.1, 0.5, 1, 2, and 5; the subsample ratio of the training instance was set to 0.75, and 1; the minimum sum of instance weights was set to 1, 2, and 3; the subsample ratio of columns was set to 0.75 and 1. Approximately 4,752 combinations of possible hyperparameters were used in XGBoost, and the root mean square error (RMSE) of each model was compared.

To prevent overfitting, we used five-fold cross-validation. Our study population was split into five groups (“five folds”) equally. In the first iteration, the first fold was used as the test dataset, and the other four folds were used as the training dataset. Using a single combination of hyperparameters, the process was repeated until all folds were used as a test dataset (five iterations). The RMSE of XGBoost was calculated by averaging the results of the five iterations, and the RMSEs of all XGBoost models were compared. The minimum RMSE of XGBoost was obtained: learning rate=0.2, maximum depth of a tree=2, gamma=2, subsample ratio of columns=0.75, minimum sum of instance weight=3, and subsample ratio of the training instance=0.75.

The optimal hyperparameters were then used to perform XGBoost to evaluate frailty with back extensor strength. The optimal XGBoost model was trained to evaluate the group with frailty (frailty score ≥3). In this study, 70% of the participants were used as the training set, and 30% of the participants were used as the test set. The cutoff value for frailty prediction was selected based on Youden’s index [36]. The performance of the optimal XGBoost model was evaluated by calculating the area under the receiver operating characteristic curve [37], with an area under the curve (AUC) of 0.5 considered as null accuracy, and a higher AUC considered as better accuracy. Other classification evaluation metrics, such as accuracy, precision, recall, and f1 score, were also evaluated to assess the predictive power of back extensor strength in the XGBoost classification.

The feature importance of each variable was estimated, and the model provided ‘gain’ as the importance of the feature in the frailty prediction tree branches [38]. The gain for each feature was calculated by dividing the sum of the gains for the feature by the sum of the gains for all the features. Features with higher gains were considered more important in XGBoost model construction. All analyses were repeated 100 times to estimate a 95% CI.

3. Results

3.1. Participant Characteristics

Of the 560 participants, 255 were male, and 305 were female (Table 1). The average age of the participants was 58.0±7.0 years, with an average of 58.5±7.0 years for men and 57.5±6.9 years for women. The muscle/fat compositions, such as TMM, BMM, PMM, AMM, VFM, and SFM, differed between men and women. Female participants had a lower waist circumference than males (p<.001). The CT results showed higher VFM in men (p<.001) and higher SFM in women (p<.001). Men had higher TMM, BMM, PMM, and AMM values (p<.001). In terms of frailty, men exhibited a faster walking speed (p<.05), higher grip strength, and higher back extensor strength (p<.001). There were no significant differences between sexes in terms of age, BMI, TFM, unintentional weight loss, self-reported exhaustion, physical activity, or Fried’s frailty score.

The participants were divided into two categories based on age to capture the impact of aging on trunk muscle/fat compositions (Table 2). There were 90 participants aged ≥65 years and 470 participants aged <65 years. The cutoff age of 65 years in this study was based on the definition of an aging society in Korea [39]. The CT results showed higher TMM, PMM (p<.05), and BMM (p<.001) in participants aged < 65 years. Also, grip strength, back extensor strength (p<.05), and walking speed (p<.001) were higher in the < 65 years group than in the ≥ 65-years group. The frailty score was higher in participants aged ≥ 65 years than in those aged < 65 years (p<.001). Waist circumference, BMI, and fat mass measurements showed no significant differences between the age groups.

3.2. Linear Regression Analysis of Trunk Muscle/Fat Compositions and the Back Extensor Strength

Table 3 shows the results of multivariate linear regression. Since age and sex had confounding effects (Table 1 and Table 2), these factors were adjusted for in the regression analysis. After adjusting for age and sex, a clear trend was observed (p<.05) in which higher back extensor strength was associated with increased AMM (r=1.12) and BMM (r=0.89). The PMM, VFM, and SFM were not significantly related to back extensor strength. Each absolute value of the VIF was <10, indicating that there was no multicollinearity issue. A relative weight analysis showed that AMM (relative weight=0.089) was the most significant predictor among the trunk muscle/fat components, and being female compared with male (relative weight=-0.118) was the most important predictor among all the parameters used.

3.3. PS Matching of the Group with the Lowest 20% Back Extensor Strength

The characteristics of the two groups were compared. One group had a lower back extensor strength of 20%, while the other group had a higher back extensor strength. Before PS matching, the age and sex covariates showed an SMD > 0.1 (Table 4). After PS matching (with a match ratio of 1:3 and a caliper of 0.01), 108 participants from the lowest 20% group and 279 participants from the higher group remained with an SMD < 0.1. Differences between the two groups were assessed using the t-test for continuous variables and the chi-square test for categorical variables. Table 4 shows that grip strength, self-reported exhaustion, and walking speed were significantly different (p<.05) between the lowest 20% back extensor strength group and the higher group, with the lowest 20% back extensor strength group showing greater frailty.

3.4. Back Extensor Strength as a New Predictor of Frailty

The results of the multivariate logistic regression showed that higher back extensor strength was significantly associated with lower odds of frailty (OR, 0.990, 95% CI, 0.983–0.997; p<.05), whereas aging was associated with increased odds of frailty (OR, 1.088, 95% CI, 1.025-1.160, p<.05). The analysis included sex and BMI as confounding factors; however, these factors were not significant predictors of frailty (Table 5).

The results of XGBoost for frailty prediction (frailty score ≥ 3 or not) showed that back extensor strength was the most important predictor of frailty (gain=0.502±0.006), and was more important than age (gain=0.325±0.005), BMI (gain=0.145±0.005), and sex (gain=0.026±0.002). The XGBoost model had an AUC of 0.579±0.004, accuracy of 0.71±0.05, precision of 0.10±0.01, recall of 0.56±0.04, and f1 score of 0.15±0.01 (Figure 2 and Table 6).

4. Discussion

The results of this study revealed a significant association between lower back extensor strength and reduced muscle volume in both the abdominal and back muscles. Additionally, participants with lower back extensor strength exhibited higher levels of frailty according to Fried’s criteria, which was characterized by weaker grip strength, slower walking speed, and more frequent feelings of exhaustion. Both the multivariate logistic regression and the XGBoost model analyses consistently demonstrated that back extensor strength was a highly significant factor of frailty, according to Fried’s criteria. The study revealed that back extensor strength’s importance in predicting frailty was greater than that of age alone, suggesting that the impact of back extensor strength on frailty outcomes outweighed the influence of age.

Previous studies have demonstrated associations between frailty and various muscle strength and volume, including the lower limb, forearm, and trunk [40,41]. However, there is a scarcity of studies that investigate the relationship between muscle strength, muscle volume, and frailty simultaneously. In this study, we evaluated both trunk muscle/fat volume and back extensor strength in the same group of participants. Through the concurrent analysis of muscle volume and muscle strength, our findings suggest a potential association between lower frailty status and increased back extensor strength, alongside increased volume in the abdominal and back muscles. In terms of rehabilitation medicine, exercise programs focused on strengthening the back extensor strength by targeting the abdominal and back muscles can be beneficial for older adults.

Descriptive analysis revealed a difference in muscle/fat composition between sexes, even when considering individuals had similar ages and BMI. Specifically, men exhibited larger visceral fat, while women demonstrated larger subcutaneous fat. These results were supported by previous studies that explained the hormonal differences between the sexes. The primary male sex steroid, androgen, stimulated muscle growth, and increased proliferation, contributing to higher levels of muscle mass in men [42]. Adipose tissues express estrogen receptors with higher activity in subcutaneous fat than in visceral fat. The elevated levels of estradiol in women act more through subcutaneous fat receptors, resulting in decreased lipolysis [43], which leads to higher SFM in women and higher VFM in men.

The multivariate linear regression analysis, which examined the relationship between trunk muscle/fat composition and back extensor strength, showed that specific parts of the trunk were associated with muscle strength. These findings align with a previous study, which also reported a significant correlation between higher back extensor strength and increased body mass and non-fat body mass, as measured using the skinfold thickness evaluation from the Durnin and Womersley method [44]. In our study, we obtained trunk muscle volume measurements using a CT scan. These specifically measured trunk muscle volumes revealed a significant linear relationship between increased back extensor strength and specific parts of the trunk, the higher abdominal and back muscles.

This study showed that the XGBoost model had low predictive power for Fried’s frailty, as measured by metrics such as AUC, accuracy, precision, recall, and f1 score. This outcome can be attributed to the multifactorial nature of Fried’s frailty criteria, including walking speed, physical activities, self-reported exhaustion, unintentional weight loss, and muscle strength. Using limited input variables such as age, sex, BMI, and back extensor strength, resulted in the limited predictive power of this XGBoost model. However, despite this limitation, the XGBoost model still outperformed the null classifier (AUC=0.5) and highlighted back extensor strength as the most influential feature among all the covariates, even more important than the age factor. These results suggest that back extensor strength may be useful in identifying individuals who are at risk of frailty.

Trunk muscle strength was assessed in this study since it has a significant effect on health outcomes. Decreased muscle volume in the lower trunk region at the mid-lumbar level has been reported to be associated with pulmonary, hepatic, and systemic dysfunctions [45]. Furthermore, researchers have emphasized back extensor strength because of its association with fall prevention in older adults, balance deficits, and other age-related symptoms. Older adults who have experienced falls tend to exhibit lower back extensor strength compared to those who have not [46]. Several studies showed that core strengthening program, which targeted the maximal isometric strength of trunk flexors, extensors, and rotators, has shown associations with delaying frailty status by achieving improvements in various aspects such as spinal mobility (maximal extension to flexion and left to right flexion of the trunk), dynamic balance (stride velocity and the Functional Reach test), and functional mobility (Timed Up and Go test) [47,48].

The trunk serves as a kinetic link that facilitates the transfer of torque and angular momentum between the upper and lower extremities during various activities [49]. This emphasizes that core strength is a significant factor in everyday performance and sports-related activities for individuals of all ages [14]. Moreover, back extensor strength has been associated not only with the risk of dependence on activities of daily living and occupational skills in daily life [50], but also with multiple age-related symptoms, such as osteoporosis, low back pain, and Parkinson’s disease [51,52]. Based on these studies, it is likely that there exist significant correlations between back extensor strength and the aging process.

Our proposition of using back extensor strength as a potential marker of frailty aligns with the concept of a vicious cycle of frailty [53]. With aging, muscles experience changes in both quality and quantity, leading to an increase in interleukin-6 and C-reactive protein levels, which stimulate inflammatory mechanisms and the aging process [54]. These inflammatory changes can lead to chronic disorders associated with systemic low-level inflammation and decreased functionality in daily life [54,55]. Since the back extensor muscles contain more muscle cells than the forearm muscles that generate grip strength, we anticipate a more pronounced decrease in resting metabolism and physical activity, potentially exacerbating the aging process.

Previous studies explored the relationship between frailty and muscle strength in different body parts, such as the connection between lower-limb muscle strength and walking speed [56]. However, this study is the first to demonstrate that not only walking speed but also self-reported exhaustion was associated with back extensor strength. This result suggests that back extensor strength can be considered a more general factor for frailty.

Our study does not propose that back extensor strength should replace grip strength as a measure of frailty. Instead, it proposes back extensor strength as an additional option for evaluating frailty. Back extensor strength can be particularly valuable when grip strength measurements are impractical, such as in cases where a cast is present on the hand. Through the combined assessment of trunk muscle volume, we have discovered the importance of trunk muscle volume in the aging process, suggesting that increasing trunk muscle strength can also be significant in preventing frailty.

Despite the widespread use of grip strength as a convenient measurement, its variability based on different postures, affecting the precision of the data, has been reported. Maximum grip strength is found in a standing posture, with the shoulder fixed forward at 45 degrees, the elbow at 90 degrees, and the wrist and forearm in a neutral position [57]. Assessing back extensor strength was done in the sitting position on a chair equipped with a portable dynamometer, allowing for improved joint stabilization, especially in the hip and knee [24,58]. This controlled position ensures reliable measurements of back extensor strength, making it a potentially reliable indicator of muscle strength, including trunk sarcopenia, similar to grip strength.

In the past, methods for measuring back extensor strength were expensive and cumbersome, which acted as a barrier to data acquisition [59]. However, recent studies have used portable dynamometers, which made back extensor strength more available in clinical settings. Moreover, the portable dynamometer provided data that were as accurate as those obtained using traditional methods [24]. As a result, we anticipate that the increased availability and accuracy of portable dynamometers will lead to more frequent analyses of back extensor strength in the future.

This study had two limitations. Firstly, the participants consisted of farmers, who are likely to have higher levels of physical activity compared to the overall population. As a result, the findings may not fully represent the general population. Secondly, frailty was defined using Fried’s criteria, which focused on specific physical components. Adopting a broader definition of frailty that incorporates overall age-related biomarkers, such as hypertension, macular degeneration, and hearing loss, as the target outcomes of the prediction model could provide a clear understanding of the relationship between aging and muscle strength. By considering a more comprehensive definition of frailty, future studies may offer valuable insights into the multifaceted nature of frailty and its potential connections to muscle strength, facilitating more holistic approaches to frailty assessment and intervention in older adults.

5. Conclusions

Our findings demonstrated a significant linear relationship between back extensor strength and the volumes of abdominal and back muscles. Furthermore, back extensor strength showed associations with multiple parameters of Fried’s frailty, positioning it as a significant indicator of frailty. The findings of this study suggest that exercise programs targeted at strengthening the abdominal and back muscles may be linked to a reduced frailty status. In conclusion, the measurement of back extensor strength could serve as an alternative tool for evaluating frailty.