Author(s): KRANZINGER, C., KRANZINGER, S.1, GLOMSER, G.2, Institution: SALZBURG RESEARCH FORSCHUNGSGESELLSCHAFT MBH, Country: AUSTRIA, Abstract-ID: 224

INTRODUCTION:
Predicting the individual effects of altitude training is of great interest in endurance sports. Estimating its potential effects results in more efficient training programme design. This study aims to investigate the possibility of predicting altitude training effects in terms of maximum increase in performance adaptation and identify key variables influencing the effectiveness of altitude training through a single performance test comprising two measurements at different altitudes using machine learning models.
METHODS:
Data of 50 athletes were analyzed who underwent the same test protocol on a cycle ergometer under two simulated altitude conditions at 400 and three hours later at 1,800 m above sea level, respectively. The altitude was simulated in a mobile altitude chamber. Physical activity was measured with a spiroergometry system from CORTEX (CORTEX Biophysik GmbH, Germany). 21 parameters (covering aspects of eg.: heart rate, respiratory rate, oxygen saturation, fat metabolism) were used for the classification approach. The most important variables were determined by applying a recursive feature elimination algorithm, using a Random Forest and taking into account the out-of-bag error (Diaz-Uriarte & de Andrés, 2005). Three classes were predicted: low, medium, and high effect of altitude training. Various statistical estimation models were tested for the classification: Random Forest, AdaBoost.M1, Support Vector Machine and Naive Bayes. The validation was carried out with a leave-two-subjects-out cross-validation.
RESULTS:
The best results were achieved with a Naive Bayes approach. The average classification accuracy across all 1,225 test data sets is 58.8% (sd: 0.34). In addition, an average Cohens kappa of 0.26 (sd: 0.49) is achieved. Examining the most important variables for classification, we find that the altitudes impact is most accurately predicted by the ratio of oxygen saturation during recovery to oxygen saturation during stress at 400 m and the ratio of respiratory rates during recovery to respiratory rates during stress at 1,800 m.
CONCLUSION:
Following the framework of Landis and Koch (1977), the results of Cohens kappa show a “fair” classification performance. According to accuracy, the expected increase in performance through altitude training can be predicted for around six out of 10 athletes. The number of test subjects with the label "high" performance improvement through altitude training should be increased, as only four of the 50 athletes have this label, which makes it difficult to classify individuals with a high expected performance improvement through altitude training.

[1] Diaz-Uriarte, R. and Alvarez de Andres, S. arxiv.org, 2005
[2] Landis, J.R. and Koch, G.G. Biometrics, 1977