Training control through artificial intelligence
A new study published in Nutrients provides a detailed overview of AI use as a training management tool. The study, authored by Gerasimos V. Grivas of the Hellenic Naval Academy and Kousar Safari of Shiraz University, explores how AI is finding its way into endurance sports - from metabolic analysis and recovery prediction to highly individualised nutrition and pacing strategies. It offers both an insight into what is possible and a sober reminder of the challenges that still need to be overcome. The message is clear: AI is changing endurance sports from general guidelines to precise performance. The study highlights AI-supported applications in the areas of metabolic monitoring, regeneration management and Personalised nutrition. AI systems are already analysing complex data sets from wearables, metabolic measurements and Training protocols. The algorithms process heart rate variability, sleep quality, nutritional data and glucose values to produce individualised recommendations for action. According to the authors, this multidimensional approach surpasses traditional coaching methods or static nutritional plans.
Researchers from the Hellenic Naval Academy and Shiraz University analysed current AI applications in endurance sports. Their analysis includes studies on marathons, triathlons and cycling, where even small physiological fluctuations can have a significant impact on the performance. Performance significantly. The scientists identified specific areas of application for artificial intelligence in sports practice.
Artificial intelligence as a forecasting tool
For decades, endurance training has been characterised by a range of well-known tools: Heart rate monitors, interval training, lactate testing, RPE scales and the athletes' own instincts. These tools were powerful, but also limited. According to the authors, they captured snapshots but not systems. They could not integrate multiple physiological streams, contextual factors or long-term trends. And they couldn't predict future states, such as the next day's recovery or energy requirements during the race. AI, Grivas and Safari argue, completely changes these dynamics. Modern AI systems take in multimodal data - heart rate variability, GPS data, sleep metrics, glucose levels, sodium levels in sweat, environmental conditions, subjective well-being, biomechanical signals - and synthesise them into patterns and predictions that even experienced coaches can't recognise. The review highlights studies in which machine learning models outperformed conventional baselines in terms of:
- Forecast of recovery the next day
- Estimation of metabolic thresholds without laboratory tests
- Optimisation of carbohydrate and fluid strategies
- Design of course-specific tempo planning
These systems not only react, but also act with foresight. They not only describe the athlete's condition, but also predict it.
The end of performance tests thanks to artificial intelligence?
According to the authors, deep learning models are suitable for estimating lactate and ventilation thresholds (VT1 and VT2) using non-invasive inputs such as heart rate, heart rate variability and power. Studies cited in the review show near clinical accuracy and offer a scalable alternative to laboratory testing. It remains to be seen whether artificial intelligence can really replace performance tests or merely interpret them better.
Optimised competition nutrition through AI
One very specific application is the control of competition nutrition. Continuous glucose monitoring devices, originally developed for diabetics, are increasingly being used by endurance athletes. AI algorithms analyse glucose levels in real time and enable precise carbohydrate intake during training and competition. This technology takes individual glycaemic reactions into account and adjusts gel timing and drink composition accordingly.
Machine learning models combine performance data and heart rate with training characteristics and environmental conditions. They estimate carbohydrate consumption and total energy expenditure during running and cycling, resulting in practical target values for carbohydrate availability and timing. However, glucose meters currently still exhibit physiological delays and device-specific errors, which is why the authors believe that AI-based nutritional recommendations should include uncertainty limits.
Recovery prediction through artificial intelligence
A twelve-week study with 43 endurance athletes demonstrated the performance of machine learning in predicting recovery. The algorithms analysed daily heart rate variability together with training load, sleep, nutrition and wellness parameters. They predicted morning recovery status and daily HRV changes more accurately than simple reference models.
AI systems now integrate additional biometric signals such as resting heart rate, respiratory rate, sleep architecture and skin temperature from wearable devices. These multimodal data sets enable more differentiated assessments of recovery trajectories. Random forest classifiers achieved accuracies between 76 and 90 per cent in fatigue detection during outdoor running.
AI models can combine measured values to predict the following:
- the perceived recovery the next morning
- Daily HRV changes
- the accumulation of fatigue
- Early warning signs of overtraining
The study refers to machine learning studies in which algorithms integrated training load, HRV, sleep duration, sleep phase distribution, nutritional quality and mood indicators. The result: prediction models that exceed simple rules of thumb and are more nuanced than commercial "recovery scores" often allow.
And data protection?
Wearables and digital platforms collect sensitive biometric and contextual data such as heart rate variability, glucose, sleep architecture and geolocation. Athletes are often given unclear information about how this data is stored, shared or misused. Consumer devices often use vague terms of use that allow the re-use of anonymised data for commercial purposes without meaningful user engagement.
Many machine learning models operate as "black boxes" and produce recommendations without comprehensible internal logic. Lack of interpretability can undermine trust, especially when model outputs contradict human expertise. Algorithm bias due to non-representative training data is another problem, as models with elite, western or male cohorts may perform less well with female, recreational or ethnically diverse athletes.
Future prospects and recommendations
According to the authors, the further development of artificial intelligence in endurance sports requires inclusive, high-quality data sets and dynamic time series approaches. They see device-aware harmonisation, rigorous external validation and close collaboration between data scientists and practitioners as a prerequisite for reliable systems. Without these measures, they believe that artificial intelligence risks being effective for the few rather than the many.
Link to the Study
Kristian Bauer
Editor
