A Machine Learning Approach to Predicting Muscle Glycogen Use During Exercise

Purpose: Here we assess the feasibility of using machine learning models to non-invasively predict muscle glycogen use during exercise. Methods: Two data sets comprised of: 1. Open-source group data from 166 studies, 2. Manually compiled granular data reported at an individual level from 8 studies published subsequently from the first data set matching the same criteria. The target variable in this study was glycogen use during exercise. Modelling was conducted on the entire data set and on four separate subsets of the data corresponding to different time durations: bands One t ≤ 20min, Two 20 < t ≤ 40min, Three 40 < t ≤ 80min, and Four t > 80min. A baseline Linear Regressor model was used for reference and four models were compared to the performance of the baseline: polynomial Support Vector Regressor (pSVR), Gradient Boosting Regressor (GBR), Random Forest Regressor (RFR) and Voting Regressor. Time band models were evaluated using 3-fold cross validation. Results: Baseline glycogen was identified as the most important variable influencing glycogen delta in all time bands. The best performing models were pSVR in band One, Voting Regressor in band Two, GBR in band Three and Linear and Voting Regressors in band Four. The magnitude of errors made by models increased with time band, from Mean Absolute Error (MAE) of 41 mmol·kg -1 in band One, to MAE of 55 mmol·kg -1 in band Four. Conclusion: We present novel predictive models for estimating glycogen utilization during exercise. Our analysis demonstrates the need for individual models for specific exercise durations.