Aerobic Fitness Level Estimation Using Wearables

Radovan Smisek1, Andrea Nemcova1, Lukas Smital1, Daniela Chlibkova2, Martin Kralik1, Jana Kolarova3, Vojtech Myska4, Martin Kolarik4, Jaromir Hubalek1
1Brno University of Technology, Faculty of Electrical Engineering and Communication, Department of Biomedical Engineering, 2Brno University of Technology, Centre of Sports Activities, 3Brno University of Technology, 4Brno University of Technology, Faculty of Electrical Engineering and Communication


Background: Aerobic fitness level (AFL) is an important indicator for everyone, especially older adults, to help them stay active. It can be useful also for athletes to improve their physical fitness.
Aims: This study aimed to assess AFL using data measured during various physical activities with wearables - accelerometric data, GPS, and photoplethysmogram (PPG) or electrocardiogram (ECG). Methods: AFL was assessed in 2D space. The first dimension is the exertion level, and the second dimension is the body’s response to the exertion. AFL is a number between 0 and 1.
Exertion level was determined based on metabolic equivalent calculated for each classified activity using the speed and elevation from GPS. Four activities were classified – walking, cycling, running, and rest. The classification is based on a combination of convolution, recurrent, and fully connected layers neural networks. The neural network inputs are data from three accelerometers (hand, ankle, chest). Body response estimation is based on heart rate calculated from ECG or PPG. Heart rate is classified into body response levels with inspiration from the well-known Borg Rating of Perceived Exertion. Results: The test set contained 31 subjects. As a reference, AFL classification by Cooper was used. It has six levels based on age, gender, weight, and the results of the Rockport Fitness Walking Test (duration and heart rate). AFL determined by our algorithm were 0.84±0.27, 0.67±0.23, 0.58±0.17, 0.45±0.16, and 0.31±0.17 for the reference classes superior, excellent, good, fair, and poor, respectively. The person with a reference condition very poor was not in our database. The correlation between the reference and determined values is 0.66. Conclusion: We have presented the method that allows AFL determination based on signals from commonly available wearable devices. This method has been tested on data measured under free-living conditions.