Athlete Monitoring: Data Analysis and Visualization
The dorem package
One of the best and the most practical papers I have read is Clarke & Skiba’s “Rationale and resources for teaching the mathematical modeling of athletic training and performance” (DOI: 10.1152/advan.00078.2011). Ever since reading that paper, I was interested in Banister’s Impulse-Response (IR) model and even wrote few R scripts myself (LINK, LINK).
Piatrikova, E., Willsmer, N., Altini, M., Jovanovic, M., Mitchell, L., Gonzalez, J., Sousa, A., & Williams, S. (Accepted/In press). Monitoring the heart rate variability responses to training loads in competitive swimmers using a smartphone application and the Banister Impulse-Response model. International Journal of Sports Physiology and Performance.
Since 2019, I have been exchanging emails with Ben Stephens Hemingway who is researching mathematical IR models and we have been talking about developing an easy-to-use R package to help fellow researchers implement various IR models. Back then, I didn’t have the necessary skills to do so, but once learning to make R packages and learning about the hardhat package I was able to create a dorem package (short of DOse REsponse Model). Hardhat is an outstanding package that aims to help developers create new model packages:
hardhat is a developer-focused package designed to ease the creation of new modeling packages, while simultaneously promoting good R modeling package standards as laid out by the set of opinionated Conventions for R Modeling Packages.
This allowed me to create a very neat and easy-to-use implementation of not only Banister IR, but also rolling average and exponential rolling average models.
Dorem also implements cross-validation and random shuffling of the predictors to test model overfit and sensitivity. My goal with dorem in the near future is to expand it to involve classification models (although it is already possible to implement it using the link function) which can be used with injury prediction models, but more importantly mixed-models to allow N>1 models (since IR models are built for N=1 or individual athletes) and estimation of the posterior probabilities of model parameters (now dorem uses optimization models that provide a single solution, but maybe in the near future I can try to implement it Stan language, but I still do not have the knowledge to do so).
Together with Ben Stephens Hemingway, Paul Swinton and Leon Greig, we are trying to develop these features. We have also submitted one review paper on the topic which is available in the preprint:
Stephens Hemingway, B., Greig, L., Jovanovic, M., & Swinton, P. (2020, September 17). A narrative review of mathematical fitness-fatigue modeling for applications in exercise science: Model dynamics, methods, limitations, and future recommendations. https://doi.org/10.31236/osf.io/ap75j
Dorem package implements a modular approach, which allows for easier extension of the package. I will use this opportunity to call developers to contribute to this project and help develop a powerful tool for fellow researchers and sports scientists.
In this video, I am explaining how the dorem package works and I am also preparing our data sets (i.e., training load and wellness) to be used in mapping training load to sleep quality.