Author(s): SCHNACK, T., SANCHEZ, R., EXEL, J., BACA, A., Institution: UNIVERSITY OF VIENNA, Country: AUSTRIA, Abstract-ID: 700

INTRODUCTION:
Heart rate (HR) is an important measure of exercise intensity in running. Common metrics, such as time-in-zone, TRIMP load, or HR kinetics, require continuous HR data.
However, even accurate ECG-based devices are susceptible to disturbed signal conduction, displacement due to motion, and incorrect placement, leading to inaccurate or incomplete HR recordings.
To reconstruct incomplete HR data, different longitudinal imputation methods exist that leverage the HR values before and after the gap for reconstruction.
This study compares the accuracy of these methods in Trailrunning. Moreover, for the first time in this context, the methods are compared across different intensity profiles and missing-data lengths.
METHODS:
Twelve amateur adult trailrunners ran three to five bouts on a 2 km outdoor trail with 165 m of elevation gain. The runners' HR and position was recorded using a Polar H10 and a Garmin GPSMAP 67i, respectively.
Missing data was artificially introduced into each bout at four locations with distinct intensity profiles: the onset of exercise, two points of constant intensity, and a point where the gradient switches from uphill to downhill. At each location, missing-data lengths of 1, 10, 25, 50, 100, 200, 400, and 800 consecutive samples were considered.
Each gap of missing data was then reconstructed using the following longitudinal imputation methods: piecewise cubic Hermite interpolating polynomial (PCHIP), linear interpolation, cubic spline interpolation, and nearest neighbor interpolation.
The root mean square error (RMSE) between methods was compared using a linear mixed-effects model, controlling for the intensity profile and missing-data length.
RESULTS:
Overall, linear interpolation showed a significantly lower RMSE than all other methods (p<0.01). The interaction between intensity profile and missing-data length revealed no condition in which linear interpolation is significantly worse than any other method.
At constant intensity, the estimated marginal means for linear interpolation are below 4 bpm for all missing-data lengths. At the onset and the gradient switch, the marginal means are below 4 bpm only for 100 missing samples or fewer. For a missing-data length of 200-800 samples, the marginal means rise above 10 bpm.
CONCLUSION:
Linear interpolation showed the best results across all conditions. Notably, it performs better than the computationally more complex PCHIP and cubic spline methods. This suggests that linear interpolation is suitable for reconstructing HR in Trailrunning. However, reconstructed HR values should be treated with caution when the intensity profile is not constant and the missing-data length is large. In these cases, RMSE values are too high for practical use because they exceed the width of typical HR zones.