Integration of a Heart Rate Prediction Model into a Personal Health Record to Support the Telerehabilitation Training of Cardiopulmonary Patients - Biomedical Engineering Systems and Technologies - page 306

2.2 Model Creation

For the integration of the predictive model into existing training systems the set of
potential predictors (input variables which explain a significant part of the response
variable) varies depending on the point in time and the use case.
We designed five different scenarios with expanding/extending datasets, which
represent typical settings of telerehabilitation training and ive training in clinics. The
first scenario describes a situation before training when the schedule is created, but no
reliable weather forecast is available (approximately three days before the training day).
The second scenario includes the weather forecast. In the third scenario the patient al-
ready wears the sensors, but the training has not yet been started. The fourth scenario
depicts an ongoing training and the prediction includes data that was gathered during
previous completed training phases. To provide an example, the average heart rate of
the warm up plateau phase can be included into the dataset for the load phase. The fifth
scenario describes the situation after the training and does also include data like the
subjective perceived exertion of the patient expressed on the Borg scale.
The following list is sorted in ascending order by the number of predictors availa-
ble and the time in relation to the training session. Each scenario expands the predic-
tor set of the previous one:
• Scenario S1 (training plan creation) : patient demographics and training plan data

(load, duration of each phase)
•

Scenario S2 (training plan creation few days before the training day): weather data

• Scenario S3 (at the beginning of the training) : resting HR, resting BP

Scenario S4 (during the training) : average HR of the former phase, HR at the end
of the former phase, BP during the load phase (phase three)
•

•

Scenario S5 (after the training) : average HR of current training phase, average HR
of load phase, average HR of all phases, recovery pulse, recovery BP, average of all
BP values, Borg value
The final list of predictors for scenario five included 24 items (see table 1).
To build a hypothesis about which values have a relevant influence on the HR, we
used a stepwise regression analysis [17]. This algorithmic approach performs a multi-
linear regression and determines a model, by adding or removing the variable with the
highest or lowest correlation of the model's F-statistics stepwisely.
So the variable with the highest chance of explaining the variance of the given
normally distributed data set is added to the model, when the correlation is big enough
to reject the null hypothesis. This is done until all variables with significant influence
(predictors) have been added and all variables with non-significant influence have
been removed from the final model. We used the standard entrance and exit toler-
ances of p ≤ 0.05 and p ≥ 0.10 for the model. Additionally, we performed chi-square
tests to confirm the normal distribution of the HR dataset.
The stepwise regression determines a set of coefficients ( B i ) and an intercept (also
called constant term) ( c ) as result. Together with a number of given predictor values ( X i )
it yields a linear combination of the following form to calculate the response variable ( Y ):
Y = c + b 1 x 1 + b 2 x 2 ... + b i x i (1)
We created such a submodel for each training phase (warm up plateau, warm up
ramp, training and cool down) to reflect the different physiological targets. These four