Biomedical Engineering Reference
In-Depth Information
Fig. 3.
Training schedule definition with integrated heart rate prediction in the PHR
Table 2 shows the accuracy of the prediction. For the calculation of average and
median over the complete training the phases are weighted by their duration.
The RMSE for scenario S1 and S2 is similar (mean ≈12.3 and median ≈11.1). This
also shows that the available weather data has nearly no effect on HR prediction. With
an average HR of ≈98.4 bpm over all training sessions, this is equivalent to a relative
mean error of ≈12.5%. The third scenario shows an average and median error of ≈8.5
and ≈6.1 which corresponds to a relative mean error of ≈8.6%. The difference be-
tween the average and mean error suggests that there are some outlier trainings that
have a strong influence on the average error.
Due to the additional predictors in S3, the median error is almost reduced to 50%
compared to S2. The main reason for this strong improvement is one dominating pre-
dictor: the resting heart rate (see table 1). The overall ranking of this predictor is dom-
inated by its S3 value of ≈41%. This strongly increases the average value where the
values are much lower in S4 (≈7.2%) and S5 (≈5.3%). This might be caused by the
dependence between resting HR and the average HR of the former phase. The latter
seems to be the better predictor.
S3 is also the scenario in which the training load has by far the highest influence
(≈5.6%) with a distance of 5% to the next smaller value in S4 (≈0.6%). A plausible
explanation for this value might be that training sessions with cardiopulmonary pa-
tients are generally conducted at a very low load of ≈35 watt on cycle ergometers.
Therefore the leg movement might have a stronger influence on the real training load,
than the selected load of the bicycle ergometer.
S4 / S5 are further increasing the precision of the prediction (mean ≈4.7 / ≈4.9,
median ≈3.2 / ≈3.5 in table 2) with an average relative error of ≈4.8% / ≈5%. This is
Search WWH ::
Custom Search