Estimating oxygen uptake in cycling using neural network analysis of easy-to-obtain inputs

AIM: Machine-learning technologies like Multi Layer Perceptron (MLP) can help to estimate physiological variables that typically require exotic hardware. For instance, direct measurement of oxygen uptake (VO2) is practically unattainable during outdoor cycling exercise. Using an Artificial Intelligence approach, the aim of our project was to predict VO2 dynamics during cycling from heart rate, power output and minute ventilation. METHODS: An MLP was used as the classifier composed of 5 fully connected layers built to predict the probability for VO2 to fall into different intensity levels (or classes). Five common classes (zones) used in the design of training programs were defined, with one additional class added for appropriate analysis (0–45% VO2max), totalling 6 classes. The dataset was formed from data collected on amateur cyclists asked to follow pseudorandom cycling on an ergometer in the laboratory. The dataset was split into a training set (that typically used in training), a validation set (that used to tune hyper-parameters) and a test set (used to assess accuracy, see Figure). RESULTS: The performance of the MLP on the test set reached 78% for the single run. The main uncertainties of the predictor were in the upper level of VO2 consumption (between classes 5 and 6), where the experimental signal showed the highest noise. Moreover, these two classes did not show a balanced representation in the training set, as such exercise intensities that elicit high VO2 are less sustainable, and more variable, above critical power. CONCLUSION: Despite modest levels of accuracy, this AI methodology holds potential to be refined and used for future cycling performance assessment from other easy to obtain variables, such as blood glucose or blood lactate concentrations.