Biomedical Engineering Reference
In-Depth Information
mainly caused by time-near HR-based values (average HR of former ≈55.7% and
current phase ≈5.6%).
Although more predictors contribute to scenario S5 a higher prediction error is cal-
culated compared to S4, whereas it was vice versa during the model building process
(≈1.56 improvement for the mean and ≈0.93 for the median RMSE). This is an indica-
tor for overfitting of the S5 model, which might occur due to the usage of too many
explanatory variables.
Table 2.
Mean error of the prediction. All values refer to the RMSE of the model in relation to
the real HR.
Scenario
Phase 1
Phase 2
Phase 3
Phase 4
Average
Median
S1
11.448 10.267 12.079 12.954 12.254 11.069
S2
11.443 10.267 12.079 12.986 12.260 11.084
S3
5.528
6.514
8.528
9.561
8.498
6.068
S4
4.347
3.636
4.637
6.572
4.733
3.266
S5
4.762
2.940
4.734
5.281
4.906
3.542
4
Discussion
The stepwise regression algorithm leads to a local optimum which is not necessarily
the global optimum. A stepwise addition of variables decreases the models' RMSE.
When using only the RMSE as an indicator for the degree of influence for each indi-
vidual predictor this has the disadvantage, that a later added predictor may have less
influence, because a part of his improvement is already explained by the previously
added variable. Thereby the result depends on the order of the steps and could lead to
a suboptimal model when applied to highly correlated variables (like systolic and
diastolic BP).
Therefore a stepwise regression can never replace expert knowledge. On a statis-
tical level, we want to improve the model by performing a factor analysis that will
reduce the number of predictors and provide a better knowledge about their correla-
tion to each other. This might also eliminate the potential overfitting of S5 and enable
the transfer to other training modalities.
The accuracy of our model strongly depends on the scenario and the associated da-
ta items. The first scenario takes place during the training plan creation and the calcu-
lated model shows the highest error. This result might still be good enough to gain an
impression about HR development of a common cardiopulmonary patient during
training time. We believe that the error of this scenario can be improved by adding
further predictors related to the patients metabolic response like weight, medication
and information about the current training state.
The available weather data only had a minor influence and lowers the precision of
the model. This may reflect the fact that weather has no direct influence on the patient
when he trains in a tempered environment. However, that does not mean that the di-
rect environment has no influence at all. We want to examine this by the measurement
of the conditions inside the training area. Furthermore we are going to examine if the
weather indirectly affects the Borg value, another very important value to control the
intensity of the rehabilitation training.
Search WWH ::
Custom Search