Construction of an anthropometric discriminant model for identification of elite swimmers: an adaptive lasso approach

Introduction

Anthropometric characteristics are widely considered as important factors affecting sport performance (Barbosa et al., 2010; Saavedra, Escalante & Rodríguez, 2010; Ugarkovic et al., 2002). Anthropometrics that matches the characteristics of a sport helps to master the movement techniques, shorten training time, and plan targeted exercise program to achieve peak performance (Putra et al., 2020). In swimming, arm length and girth, or the ratios (e.g., hand-length/forearm-length and forearm-girth/wrist-girth) were found to be significantly associated with the performance of breaststroke (Sammoud et al., 2018a), front crawl (Jürimäe et al., 2007; Lätt et al., 2010) or freestyle (Nevill, Oxford & Duncan, 2015), backstroke (Sammoud et al., 2019), and butterfly (Sammoud et al., 2018b), highlighting the limb advantage of longer levers. In addition, leg length, trunk breadth (Sammoud et al., 2018a), and sitting height (Sammoud et al., 2019) were also found to affect swimming performance, but their effect could vary by stroke. From these findings, it is reasonable to hypothesize that anthropometric factors could be useful in the identification of elite swimmers.

Indeed, previous studies have examined the feasibility of identifying elite swimmers by utilizing anthropometric factors. However, in these studies, ordinary regression and discriminant analysis were commonly used to explore the relationship between anthropometric factors and swimming performance (Lätt et al., 2009a, 2009b; Saavedra, Escalante & Rodríguez, 2010). The sample sizes are often very limited, with a large number of different anthropometric factors available, and traditional methods often perform poorly. In this study, we tackled the relationship between multiple anthropometric factors and swimming performance using an adaptive Lasso (Least Absolute Shrinkage and Selection Operator) method (Bhattacharya & McNicholas, 2014; Schneider & Wagner, 2012), which has advantages in consistent variable selection. Thus, the main aim of this study is to build a discriminant model using anthropometric factors to identify elite swimmers through an adaptive Lasso approach.

Materials and Methods

Results

Table 1 illustrates the basic information of the swimmers. A total of 71 continuous variables composed a 71 × 71 lower triangular matrix with the diagonal and the above elements being zero. The pairwise correlation analysis eliminated 48 continuous variables, leading to a total of 24 candidate variables (23 continuous plus gender) for subsequent analyses (Table S3).

In the whole-sample model, gender was firstly converted into a dummy variable, with 0 being male and 1 being female. The obtained coefficient solution path and the corresponding BIC were shown in Figs. 1 and 2, respectively, and the minimum BIC was −465.387. The penalty parameter was then chosen to shrink the coefficients of some of the candidate variables to zero (i.e., excluded from the model), and finally 12 out of the 24 variables were selected (Table 2).

According to the chosen variables, a logistic model was built and statistical inference was performed (Fig. 3). The predicted values were between 0 and 1, representing the probability of becoming an elite athlete. The estimated coefficients of the selected variables were shown in Table 2. This model explained 48.23% of the deviance. The receiver operating characteristics (ROC) curve was shown in Fig. 4, with an AUC (obtained by the bootstrapping) of 0.926 (95% CI [0.895−0.956]; P = 2.42 × 10⁻²⁹).

In stratified analysis by gender, nine out of 23 variables were selected for male swimmers (Table 3), while eight variables for female athletes (Table 4). The model for male explained 46.49% of the deviance, with an AUC of 0.921 (95% CI [0.880–0.963]; P = 8.82 × 10⁻¹⁷; Fig. 5). The model for female explained 54.85% of the deviance, with an AUC of 0.941 (95% CI [0.898–0.984]; P = 7.67 × 10⁻¹⁵; Fig. 6).

Discussion

In this study, we built a mathematical model using anthropometric factors to identify elite freestyle swimmers (≤400 m) through an adaptive Lasso approach. The adaptive Lasso selected 12 variables using the whole sample, and six of them were consistently selected in sex-specific analysis. The models showed satisfactory performance in the identification of elite short-to-medium-distance freestyle swimmers. To the best of our knowledge, this is the first study that adopted the adaptive Lasso approach to identify elite swimmers using anthropometric characteristics.

Fat free mass, which primarily consists of skeletal muscle, was reported to be one of the best predictors of swimming performance (Lätt et al., 2009a; Nevill, Oxford & Duncan, 2015). For example, freestyle swimming speed could increase by 33.1% with an increase of 1-kg lean body mass (Nevill, Oxford & Duncan, 2015). More lean body mass often means more appendicular lean mass as well as more skeletal mass in the limbs, leading to greater stroke rate and subsequent propulsive force, and finally quicker swimming speed (Nevill, Oxford & Duncan, 2015; Sammoud et al., 2018b). Consistent with these findings, in our study, skeletal muscle mass was selected as a significant predictor of elite swimmers: with 1-kg increase in skeletal muscle mass, the odds of becoming an elite swimmer increased by 33% (OR = 1.33, 95% CI [1.17–1.54]; P < 0.0001) and 67% (OR = 1.67, 95% CI [1.24–2.43]; P = 0.002) in the analysis using the whole sample and female swimmers, respectively (Tables 2 and 4). Although skeletal muscle mass was not significantly associated with being an male elite swimmers, it still contributed to the prediction accuracy (OR = 1.19, 95% CI [1.00–1.44]; Table 3). These findings corroborated the importance of muscularity component in performance prediction and talent identification.

The trunk is an essential segment for swimming performance. Long torso was found to positively affect backstroke swimming speed (Sammoud et al., 2019). Expanding to the overall competitive performance (i.e., not restricted to one stroke and/or distance), torso length (sitting height) was the only anthropometric variable selected in the multiple lineal regression model in male swimmers, and sitting height explained 34% of the variance in swimming performance (Saavedra, Escalante & Rodríguez, 2010). In our study, we found that torso length contributed to the prediction of being an elite swimmer: with 1-cm increase in torso length, the odds of becoming an elite swimmer increased by 34% (OR = 1.34, 95% CI [1.15–1.59]; P < 0.0001). Stratification analysis by gender yielded similar results (Tables 3 and 4). These results were further supported by previous findings regarding the advantage of having shorter legs in freestyle swimming, as wave drag could be reduced due to longer torso, leading to increased hull speed and decreased energy consumption (Nevill, Oxford & Duncan, 2015; Papic & Sanders, 2018; Sammoud et al., 2019). Long torso is thus a promising indicator of elite swimmers.

In addition to length of the trunk, horizontal size of the trunk also plays an important role in swimming performance. Biacromial breadth (r = −0.61, P < 0.01) and biiliocristal breadth (r = −0.46, P < 0.01) were significantly correlated with swimming sprint performance in young male swimmers (Geladas, Nassis & Pavlicevic, 2005). Furthermore, each 1-cm increase in biacromial breadth and biiliocristal breadth increased 100-m butterfly speed by 48.9% and 29.2%, and 100-m breaststroke speed by 56.5% and 40.3%, respectively (Sammoud et al., 2018a, 2018b). In our study, trunk breadth ratio (i.e., bitrochanteric breadth/biacromial breadth) and trunk circumference ratios (i.e., waist circumference/chest circumference, and waist circumference/hip circumference) were selected as significant predictors: in the whole sample, a 1-unit increase in the first two ratios decreased the odds of being an elite swimmer by 18.2% (OR = 0.818, 95% CI [0.698–0.947]; P = 0.009) and 23.2% (OR = 0.768, 95% CI [0.647–0.897]; P = 0.001), respectively; while a 1-unit increase in the last ratio increased the odds by 27% (OR = 1.27, 95% CI [1.107–1.474]; P = 0.001) (Table 2). Similar results for trunk circumference ratios were observed in stratified analysis by gender (Tables 3 and 4). These observations suggested that greater biacromial breadth and chest circumference were associated with higher likelihood of being an elite swimmer, while bitrochanteric breadth and hip circumference showed a reverse association with the likelihood of being an elite swimmer. Our results, combined with the negative association of head circumference, supported that a ‘streamlined’ body shape was essential to an elite swimmer as it could reduce resistance during the underwater phase (Sammoud et al., 2018b). Together, our findings suggested that coaches should take trunk into account as a key anthropometric predictor of being an elite swimmer.

We found different indicators for male and female elite swimmers, implying that there may be differences in anthropometric characteristics that affect swimming performance in male and female swimmers. Previous literature attributed gender differences in swimming performance to differences in somatic features (Knechtle et al., 2020). For example, having a higher lean body mass and lower body fat in male swimmers compared to females was found to be significantly associated with shoulder strength and swimming performance (Avlonitou et al., 1997). In addition, anthropometric characteristics also contributed to differences in swimming biomechanics (e.g., upper and lower segment coordination, energy cost, and swimming economy) between male and female swimmers. For instance, female swimmers have a smaller body size, which allows them to have lower body torque due to reduced distance between the center of mass and the center of volume, leading to a smaller body cross-sectional area in the water and improved swimming economy (Seifert, Barbosa & Kjendlie, 2011).

Early (de)selection and early specialization is a concern for long-term athlete development, while early diversification is considered as a remedy measure for early specialization. Both early age of peak performance of swimmers and the “10-year rule” suggested the need for early specialization of swimmers (Born et al., 2022). However, the unique performance environment of swimming limits effectiveness of early diversification practices in transferring performance elements and cross-training from other diversified sports (Baker, 2003). Nonetheless, dry-land strength training programs are still important for swimmers as upper-body strength was found to contribute significantly to the full stroke cycle: with 1-kg increase in the upper-body dry-land strength, the thrust is increased by 0.397 N (95% CI [0.189–0.605]; P = 0.002) (Morais et al., 2020a).

This study has some limitations. We only considered anthropometric factors in the model, and the included somatic indicators were not comprehensive. Future studies are needed to explore whether adding other factors (e.g., physiological and bioenergetical factors) and other anthropometric factors (e.g., frontal surface area, trunk transverse surface area, maximal rate of change in cross-sectional area, and posterior form gradient of the waist-hip) that indicate actual athletic swimming performance can further improve the performance of the prediction model (Morais et al., 2011; Morais et al., 2020b; Papic et al., 2020). In addition, considering test efficiency, we only used the traditional soft tape to collect circumference features, while several studies reported that new procedures, such as 3D scanners, could provide reliable and accurate data (Rumbo-Rodríguez et al., 2021). We only used cross-sectional data in model building. It is possible that changes in anthropometric factors over time may be better indicators for identification of elite swimmers (Abbott & Collins, 2002). Future studies utilizing longitudinal data and advanced analytical methods, such as joint modeling (Wulfsohn & Tsiatis, 1997) and time-varying effects modeling (Tan et al., 2012), might provide better predictive performance. Joint modeling takes both longitudinal follow-up data and time-to-event information into account, leading to more precise estimation and more accurate prediction than modeling the two types of data separately (Koslovsky et al., 2020; Li & Luo, 2019; Rizopoulos et al., 2016), while time-varying effects modeling targets intensive longitudinal data, and can provide dynamic predictions when more longitudinal data are available, making it ideal for identification of elite swimmers. To represent the homogeneity of swimmers as much as possible, event specificity was not strictly controlled for in this study. Future studies are needed to explore anthropometric factors associated with each specific event by recruiting swimmers who participate in the same events after sample size calculation. In addition, our study is of retrospective nature. Prospective studies that align sampling time with the time at which an athlete achieves his/her highest level could provide more robust and accurate prediction. This study was conducted on Chinese swimmers, and it is not known whether our findings can be generalized to swimmers of other ethnicities. More studies with data from other ethnicities are needed to validate our findings.

Conclusions

The present study adopted the adaptive Lasso method to select anthropometric factors to establish a model for the prediction of elite short-to-medium-distance freestyle swimmers. Our analysis pinpointed several anthropometric factors that were associated with being an elite swimmer, such as torso length, leg length A to H, head circumference, ratio of waist to hip, and skeletal muscle mass. The models showed satisfactory predictive performance, implying that they might be helpful to coaches in the identification of elite freestyle swimmers in the late adolescence. They also indicated that the adaptive Lasso approach can be applied in talent identification in other sports.

Supplemental Information