An intelligent strategy for endurance training based on a virtual lactate sensor

Abstract

In this thesis, a first fully operational virtual LT sensor was created for recreational runners. This way, a so demanded operational solution to help the training of recreational runners was created. Moreover, the Lactatus software was created to guide, ease the athletes' LT estimation process and implement the additional information obtained in this thesis into their training decision-making process. This way, the work of this thesis is made tangible, widely available and usable to recreational runners. This solution grew from the creation and formalization of a strategy to help pose and apply ML to complex phenomena, an important contribution of this thesis. This strategy combined an iterative meta-process and a satisficing approach to deal with the problem boundary discovery and reduce the problem complexity. Then, the design of the virtual LT sensor was divided into three steps: context characterization, content representation and next step decision. The formalization of this methodology and a modification of next step decision are novel contributions. Additionally, several novel techniques are used, including a standardization of the temporal axis, a modified stratified sampling and a computational algorithm to discover the inherent noise that the features may contain. This way, a robust strategy and methodology is created to design virtual sensors for problems with similar characteristics. The application of this methodology led to an important conclusion. Concretely, the Dmax LT intrinsic error analysis showed that a higher accuracy of the virtual LT sensor was unnecessary and even non-characterizable. This manifested the importance of understanding the variability of the output features with respect to the input errors. The computation algorithm also allows to LT protocols could also be evaluated from this perspective in order to quantitatively address their reliability. This may allow to make an objective cross-comparison of the accuracy of different LT protocols, something that, is not well addressed in the literature. One of the possible limitations of this solution is that the recreational runner population here characterized may not be representative of recreational runners of other culture, ethnicity or different contexts. However, one of the main advantages of providing a simple solution is that, unlike other black-box models, it is easily reproducible and adjustable, meaning that we have set a common ground for other researchers to evaluate the impact of our proposal. In the best-case scenario, future experiments done in other contexts will validate that we have been capable of discovering a common characteristic of recreational runner population. In the worst-case scenario, we have provided an easy to follow methodology and a strong prior that will allow to adjust the estimator according to individual characteristics of different populations.