Barsumyan Artur, Shyla Raman, Saukkonen Anton, Soost Christian, Graw Jan Adriaan, Burchard Rene
Faculty of Medicine, Philipps-University of Marburg, Marburg, Germany.
Sports Medicine and Joint Centre, Department of Orthopedics and Trauma Surgery, Lahn-Dill-Kliniken, Dillenburg, Germany.
Front Artif Intell. 2025 Jul 4;8:1623384. doi: 10.3389/frai.2025.1623384. eCollection 2025.
The most important parameter influencing performance in endurance sports is aerobic fitness, the quality of the cardiovascular system for efficient oxygen supply of working muscles to produce mechanical work. Each individual athlete responds differently to training. However, for coaches it is not always easy to see improvement, accumulated fatigue, or overreaching. In the new era of technology, we propose an experimental method using machine learning (ML) to measure response quantified as aerobic fitness level based on cardiovascular drift and aerobic decoupling data.
Twenty well-trained athletes in cycling-based sports performed monthly aerobic fitness tests over five months, riding at 75% of their functional threshold power for 60 min. Based on aerobic decoupling (power-to-heart rate ratio) and cardiovascular drift of each test ride, a prediction model was created using ML (Logistic regression, Variational Gaussian Process models and k-nearest neighbors algorithm) that indicated whether or not an athlete was responding to the training. Athletes were spitted as responders (i.e., those showing improvements in cardiovascular drift and aerobic decoupling) or non-responders.
Cardiovascular drift and aerobic decoupling demonstrated a significant strong linear correlation. All ML models achieved good predictive performance in classifying athletes as responders or non-responders, with cross-validation accuracy ranging from 0.87 to 0.9. Average predictive accuracy of 0.86 was for k-nearest neighbors, 0.91 for logistic regression, 0.93 for Variational Gaussian Process model. The Variational Gaussian Process model achieved the highest classification for training response.
Cardiovascular drift and aerobic decoupling are reliable indicators of response to training stimulus. ML is a promising tool for monitoring training response in endurance sports, offering early and sensitive insights into fitness adaptations or fatigue that can support more personalized training decisions for coaches and athletes.
影响耐力运动表现的最重要参数是有氧适能，即心血管系统为工作肌肉有效供应氧气以产生机械功的能力。每个运动员对训练的反应各不相同。然而，对于教练来说，要看出运动员的进步、累积疲劳或过度训练并不总是那么容易。在技术新时代，我们提出一种使用机器学习（ML）的实验方法，根据心血管漂移和有氧解耦数据来测量量化为有氧适能水平的反应。
20名从事自行车项目的训练有素的运动员在五个月内每月进行一次有氧适能测试，以其功能阈值功率的75%骑行60分钟。基于每次测试骑行的有氧解耦（功率与心率比）和心血管漂移，使用ML（逻辑回归、变分高斯过程模型和k近邻算法）创建了一个预测模型，该模型可表明运动员是否对训练有反应。运动员被分为反应者（即心血管漂移和有氧解耦有改善的运动员）或无反应者。
心血管漂移和有氧解耦呈现出显著的强线性相关性。所有ML模型在将运动员分类为反应者或无反应者方面都取得了良好的预测性能，交叉验证准确率在0.87至0.9之间。k近邻的平均预测准确率为0.86，逻辑回归为0.91，变分高斯过程模型为0.93。变分高斯过程模型在训练反应分类方面取得了最高准确率。
心血管漂移和有氧解耦是对训练刺激反应的可靠指标。ML是监测耐力运动训练反应的一个有前途的工具，可为体能适应或疲劳提供早期和敏感的见解，从而支持教练和运动员做出更个性化的训练决策。
