Author(s): STESSENS, L., MEEUSEN, R., AERTS, J.M., Institution: KU LEUVEN , Country: BELGIUM, Abstract-ID: 1785

INTRODUCTION:
The lactate threshold (LT) is an important marker in endurance sports used to identify training zones and monitor training progress. The determination of the LT requires expensive equipment, invasive sampling and a visit to a specialized facility. Therefore, regular monitoring of the LT is inaccessible to a large population of cyclists, even among the elite. Mathematical modelling might form a suitable method for continuous monitoring of performance, especially in combination with commercial sensors and wearable technology. In cycling, the heart rate, power and cadence are already continuously measured and monitored with commercial sensors during training and competition, making them particularly suitable for integration in a modelling technique. This study attempts to explore the estimation of LT with linear time-invariant and linear time-varying models solely based on the heart rate and power.
METHODS:
11 amateur, trained cyclists (6 male, 5 female) participated in this study. They performed an incremental performance test on a Lode Excalibur ergometer in laboratory conditions to determine the maximal oxygen consumption (VO2max), ventilatory thresholds (VT) and LT. The workload increased with 40 watts every 3 minutes. The heart rate was measured with a 12-lead electrocardiogram (ECG). Modelling and analysis was performed with the CAPTAIN toolbox in Matlab R2021b. The LT was estimated using a discrete-time transfer function approach, which was selected due to the low computational effort required and robust qualities. The results were compared with the actual LT2, calculated with the modified Dmax method. Linear time-invariant parameter (TIP) and time-varying parameter (TVP) models were compared.
RESULTS:
Calculation of the LT with TIP models was performed with an average error of 11%. For 5 out of 11 participants, the estimated LT was approximated with an error smaller than 10 watts. TVP models performed with an average error of 4%. For 9 out of 11 participants, the LT was estimated with an error smaller than 10 watts. Since the heart rate response is a non-linear signal, a linear TIP model is not able to capture the variability of the signal. By introducing time-varying parameters this non-linearity can be accounted for, even with a linear model.
CONCLUSION:
Modelling techniques based on the heart rate and power output approximate the LT with a decent accuracy, with TVP models performing better than TIP models. Our results are interesting to the ECSS community since they propose a more accessible complement to the current gold-standard of testing that might enable regular, day-to-day monitoring of the LT, as well as open up up LT testing to a wider audience. Given the widespread popularity of heart rate, power and cadence sensors in the cycling population this modelling approach could even be applied in the field. Future work needs to focus on verifying the technique over a longitudinal time period and adapting it to field-captured data.