Predicting Cycling Performance Before and After Training: Insights From Machine Learning Using Small Samples

Research on optimizing sports performance is challenging due to large individual variability and small sample sizes. Machine learning (ML) is underexplored in predicting performance changes following training interventions. This study aimed to predict 4-km cycling performance before and after a 12-week intervention using ML models, incorporating physiological, training load, and well-being data while identifying key predictors. Twenty-seven recreational cyclists completed baseline tests, including 4-km time-trial, {V.O2max, pulmonary {V.O2-kinetics, Wingate, squat jump, ultrasound imaging, and anthropometry. They were randomly assigned to one of four training programs. ML models (generalized linear models (glm), random forest (rf), and principal component regression (pcr)) were used to predict performance pre- and post-training, as well as performance changes. Models were evaluated using R² and mean absolute error (MAE). Cyclists produced 4.1 ± 0.7 W/kg during the time-trial. Change in cycling performance showed substantial individual variability and did not differ between training programs (p > .05). ML models accurately predicted performance pre-training (R² = 0.875, MAE = 0.260 W/kg) and post-training (R² = 0.792, MAE = 0.266 W/kg) using glm, but changes in performance were less predictable. Key predictors included power at {V.O2max, ventilatory thresholds, body composition, deoxygenation, sleep, and sickness. Findings highlight ML’s potential for predicting endurance performance but indicate difficulty in forecasting individual adaptations to training. Machine learning models enabled accurate predictions of time-trial performance in recreational cyclists using a comprehensive set of predictors, although changes in performance were more difficult to predict.Accurate predictions could be obtained even with a small sample size, while showing high generalizability using internal validation, after applying techniques to reduce overfitting.Key predictors of time-trial performance include physiological determinants of endurance performance, parameters related to body composition, training impulse, sleep, sickness, and well-being. Machine learning models enabled accurate predictions of time-trial performance in recreational cyclists using a comprehensive set of predictors, although changes in performance were more difficult to predict. Accurate predictions could be obtained even with a small sample size, while showing high generalizability using internal validation, after applying techniques to reduce overfitting. Key predictors of time-trial performance include physiological determinants of endurance performance, parameters related to body composition, training impulse, sleep, sickness, and well-being.