Comparing methods for summarizing a training load in prediction models of swimming performance
Résumé
Introduction. Training quantifications are valuable for monitoring and prescribing elite swimmers’ training and are indispensable in mathematical models that attempt to accurately predict performance. Modelling the association of training with performance raises an important issue: how should we account for volumes at different training intensities? Mujika et al. (1996) constructed a training load by adding weighted (by a priori constants representing energetic intensities) volumes from each intensity. Avalos et al. (2003) computed a training load as the sum of normalised training intensities. Here we compared the predictive accuracy of these methods to others based on: a/ alternative normalisations, b/ summary scores derived from data, and c/ machine learning techniques, with recognised predictive qualities, such as PLS.
Methods. Training volumes at eight intensity levels (in kilometres and minutes per week, for in-water and dry-land workouts, respectively) and performances in competition of 138 professional French swimmers were collected during 20 seasons. Training intensities were determined using measurements of blood lactate concentrations. We assumed that swimmers may react differently to the same training and over time, thus we used mixed-effects models adjusted for sex, age, swimming distance and event specialty. The comparison criterion was the cross-validated prediction error.
Results. Summary scores for three training loads (low-intensity/high-intensity/dry-land workouts) with data derived weights showed the best results (mean cross-validated prediction error ± SD were 0.60±0.89, 0.50±0.62 and 0.10±0.19 for sprint, mid- and long-distances, respectively). However, cross-validated prediction errors were close relative to their variances, which were high.
Conclusions. The use of complex machine learning techniques did not lead to more accuracy in predicting performance. Although data derived scores showed the lowest prediction error, the statistical variability was too high for being conclusive. A possible explanation is that the lactate sensitivity to extraneous factors (mode of exercise, technique quality of training, diet or sleep quality prior to test) and the subject-specific variations in lactate thresholds introduce not negligible measurement error. As practical recommendation, we suggest completing lactate measurements with athlete/coach questionnaires to better assess the physiological stress associated with the training load. Also, errors-in-variables models might be more appropriated.