Rapid and cost-efficient assessment of CHO consumption among endurance athletes is challenging on the field. Although multiple dietary assessment tools, such as FFQs, 24 h recalls and dietary journals, are available to calculate an athlete’s CHO intake, these tools usually take a lot of time to complete and require the experience of a trained professional for analysis. This, combined with the fact that large proportions of endurance athletes do not meet the recommended CHO intake, is a concerning issue. Here, we have developed a CHO-specific dietary screener that allows rapid detection of endurance athletes at risk of not achieving a CHO intake of 6 g/kg of BW or more. To our knowledge, this is the first validated tool that screens for adequate CHO intake among athletes.
The final model upon which the screener is based has both a high sensitivity and specificity in the target population (89.5 and 87.3%, respectively), which are desired traits . Such statistics indicate that the screener is as accurate in adequately identifying athletes who meet and those who do not meet the recommendation for CHO intake. The high AUC of the ROC curve (or c-statistic) yielded by the 15-variables model (0.94 on a range from 0.5 to 1.0) is also reflective of a dietary screener that has excellent accuracy . Furthermore, the model’s NPV was considerably higher than its PPV (94.5% vs. 77.3%), indicating that the screener is slightly more accurate in identifying endurance athletes who do not achieve adequate CHO intake than those who do. Such characteristic is highly desirable in the context of this research, as the ultimate goal of the CHO screener is to target athletes who would benefit from nutritional counseling, i.e. those with inadequate CHO intakes.
Very few studies have used an approach similar to ours to develop predictive models of adequate/inadequate dietary intakes, which makes comparison difficult. Most attempts were undertaken with a health rather than sports perspective. In those previous studies, predictive models and tools often achieved either a high sensitivity or a high specificity, but rarely both. For instance, Cook et al. built single-question and five-question screeners to rapidly assess fruits and vegetables intake among non-athletes. Sensitivity values ranged from 35.7 to 45.5% while specificity values ranged from of 81.8 to 84.9% among all five-question screeners developed by the research team. Using a single-question approach yielded high sensitivity but low specificity in the same population . In most of these studies, the AUC of the ROC curves were fairly low, indicative of poor accuracy.
In an attempt to develop the simplest and yet most accurate CHO screener possible, we gave important considerations to limitations specific to the sports work environment. First, we had access to numerous dietary variables for the development of the model, such as energy, vitamin and protein intake, which may have contributed to a better prediction accuracy. However, such information is not readily available to either the respondent or the resource responsible for the screening test. It was therefore decided a priori to exclude such information. All anthropometric measures were also a priori excluded as they are too-closely related to the outcome measure to predict, which is based on BW. Similarly, cut-off values for each predictive food in the model were rounded to full daily or weekly servings to facilitate screener administration.
Several methods can be used to develop the predictive model of an outcome. Here, a multifaceted approach was used, but ultimately a stepwise logistic regression modeling approach yielded the final model. A classification tree (CT) approach was also considered to develop the screener. This method uses discriminant analysis to test various combinations of variables in order to maximize the CT’s predictive power . Different algorithms can be used to build CTs; the CART algorithm was the chosen method for our purpose. What characterizes the CART algorithm is that it builds a very large CT and then prunes it to a smaller size to minimize classification errors . A 10-fold cross-validation is used to prune the initial CT. The use of this method would have been beneficial for this particular research since ideal cut-off points are calculated directly in the CT algorithm. Unfortunately, this method yielded underwhelming results, with unacceptably high values for false negatives (approximately 30%) when applied to the athletes sample (the VALID cohort). We hypothesize that the sample of non-athletes used to develop the CT may have comprised too few individuals with a CHO intake > 6 g/kg of BW, thereby reducing the data usable by the algorithm to maximize the CT’s predictive power.
Although this is the first study to develop a CHO-specific dietary screener for endurance athletes, limitations should be noted. First and foremost, the sample used to build the screener for application among athletes comprised non-athletes. This may have been a very significant shortcoming, considering that the diets of non-athletes and of endurance athletes are quite different. Second, a small proportion of the sample of individuals used to develop the screener achieved an intake of CHO greater than 6 g/kg of BW, which may have hindered our ability to accurately predict this nutritional outcome. Ideally, the development of this CHO-specific screener would have been based on data from a large cohort of endurance athletes, but this was not possible. Third, the target of 6 g CHO/kg BW may not be applicable to every endurance sport or training regimen, and this is a limitation when using the screener among athletes whose CHO needs are greater than 6 g/kg of BW. Furthermore, participants in the development sample were not asked about CHO supplements often used in endurance sports. This is a limitation of the screener as intake of CHO is influenced by the use of such supplements in athletes. Nevertheless, the accuracy and hence validity of the CHO-specific screener among endurance athletes is considered to be excellent, despite these limitations. Lastly, exploring different approaches for model development is a strength considering that very few studies in the field of nutrition have used CTs to create predictive models.