Estimating an individual’s oxygen uptake during cycling exercise with a recurrent neural network trained from easy-to-obtain inputs: A pilot study

Measurement of oxygen uptake during exercise (V˙O2) is currently non-accessible to most individuals without expensive and invasive equipment. The goal of this pilot study was to estimate cycling V˙O2 from easy-to-obtain inputs, such as heart rate, mechanical power output, cadence and respiratory frequency. To this end, a recurrent neural network was trained from laboratory cycling data to predict V˙O2 values. Data were collected on 7 amateur cyclists during a graded exercise test, two arbitrary protocols (Prot-1 and -2) and an “all-out” Wingate test. In Trial-1, a neural network was trained with data from a graded exercise test, Prot-1 and Wingate, before being tested against Prot-2. In Trial-2, a neural network was trained using data from the graded exercise test, Prot-1 and 2, before being tested against the Wingate test. Two analytical models (Models 1 and 2) were used to compare the predictive performance of the neural network. Predictive performance of the neural network was high during both Trial-1 (MAE = 229(35) mlO2min-1, r = 0.94) and Trial-2 (MAE = 304(150) mlO2min-1, r = 0.89). As expected, the predictive ability of Models 1 and 2 deteriorated from Trial-1 to Trial-2. Results suggest that recurrent neural networks have the potential to predict the individual V˙O2 response from easy-to-obtain inputs across a wide range of cycling intensities.


Introduction 1 Content
Relevant output figures are reported here from the application of the neural network model and of the analytical model (model 1 and 2). Personal, anthropometric and physiological data of the participants are also provided.   for Trial 1 and 2, where Protocol I and Wingate tests were used to validate the models, respectively.

Sample data
Detailed information regarding the composition of our sample is provided in tab. 2.

Individual responses
Individual responses are provided in this section. Given that the plots are automatically generated, information regarding the number of the participant and trial1/2 settings are provided in the title and in text boxes.       Figure 10: The autocorrelation plot (correlogram) is used to assess the correlation between the signal and the delayed (lagged) replication. Sample autocorrelation (blue line) is reported versus time lags. The horizontal (grey) lines displayed in the plot correspond to 95% and 99% (dashed) confidence bands. Figure 11: Training set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red), cadence (ω, green) andV O2 (VO2, blue). Every quantity is normalised on the maximal value reported in the individual .json file. Figure 12: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.     Figure 18: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.     Figure 24: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.     Figure 30: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.     Figure 36: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.                                   Figure 72: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.           Figure 84: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.

Adding a new participant
In this section we report the results obtained by training the neural network with just an incremental-toexhaustion test. This participant was not included in the main study. Here, we want to check whether the number of test needed to train the neural network can be reduced to a single test. This should reduce the time spent by the athletes in the laboratory. We tested the neural network on two different high-intensity interval training protocols.
The neural network can be trained and tested with a new participant as follows: (a) Collect and save the data in .csv files.
(b) Create a .json file and indicate maximal values (e.g. those obtained during the incremental to exhaustion test).
(c) Modify the code and include the input/output that you would like to use (e.g. you have other info like oxygen saturation or you would like to use RF as output).
(d) Train the neural network.
(e) Test the neural network.

Test a new participant
The results of the predictions (i.e. residuals, correlation, autocorrelation and Bland-Altman) and the plots of the training/testing datasets are provided in the following figures.     Figure 90: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.     Figure 96: Test set include: power output (P, black), respiratory frequency (Rf, gold), heart rate (HR, red) and cadence (ω, green). TheV O2 value is assumed to be unknown during the test set, therefore is left undefined in this plot. Every quantity is normalised on the maximal value reported in the individual .json file.

Analytical models
In this section we report the results obtained during the validation of the analytical models (model 1 and 2). Summary statistics such as: mean absolute error ( ), mean absolute percentage error ( %), variance explained R 2 and root mean square error (RMSE), are also reported.