INTRODUCTION: Coaches design training programs consisting of a variety of training types to optimally prepare their athletes. However, previous studies showed a possible misalignment between coaches` intentions and athletes` physiological reactions to training session. To investigate this misalignment, Machine Learning (ML) techniques were used to compare physiological responses between different training types, possibly revealing unique training type characteristics and allowing coaches to optimize their training programs design. Therefore, training types were predicted by classification models using internal and external training load parameters. METHODS: Heart rate data, lap times and session Rating of Perceived Exertion (sRPE) were collected from 17 elite short track speed skaters during the 2019-2020 season. A Kruskal-Wallis test with Dunn post hoc test was performed to analyze differences between the short track speed skating specific TRaining IMPulse Short Track (TRIMPST) and the speed zone-based Speed Score between training types. Using TRIMPST and Speed Score along additional training load parameters, a K-Nearest Neighbors (KNN), Decision Tree Classifier (DTC), Random Forest Classifier (RFC) and Logistic Regression (LR) were computed. These classification models were used to predict training type using internal and external training load parameters. The models were evaluated using accuracy, precision, recall and F1 scores. RESULTS: A Kruskal-Wallis test showed significant differences in TRIMPST and Speed Score between training types, 2(6) = 86,37, p<0.001 and 2(6) = 126,71, p<0.001, respectively. Accuracy scores for the classification models for external load were 0.75, 0.75, 0.71 and 0.49 for the KNN, RFC, DTC and LR models, respectively. Accuracy scores for the classification models for internal load were 0.37, 0.44, 0.38 and 0.46 for the KNN, RFC, DTC and LR models, respectively. The classification models using external load parameters performed well in predicting extensive interval (EXT INT) and intensive interval (INT INT) trainings, whereas the classification models using internal load parameters performed well in predicting extensive endurance (EXT END). CONCLUSION: Our approach opens a novel perspective on training program evaluation and training load monitoring. Using external training load parameters, training types can be accurately predicted. However, our findings show that training types cannot be accurately predicted using internal load parameters. Furthermore, since there was high internal intraclass variance in the internal load parameters, physiological responses to training load appear to be highly individual. Since the internal and external load classification models are unable to predict training types with accuracy, there appears to be a misalignment between coaches and athletes. It appears more intense training sessions cannot be accurately predicted using internal load measures.
© Copyright 2022 27th Annual Congress of the European College of Sport Science (ECSS), Sevilla, 30. Aug - 2. Sep 2022. Veröffentlicht von Faculty of Sport Science - Universidad Pablo de Olavide. Alle Rechte vorbehalten.