Biomedical Engineering – From Theory to Applications

54
Biomedical Engineering – From Theory to Applications
The non linear solution accounted for 36% of WL variance, significantly higher than 10% of
the linear model using the same independent variables: this indicated a better fit for the non
linear model.
Furthermore, subjects were assigned to different groups according to actual WL quartiles in
order to evaluate the classification (ROC curves) and prediction (cross-validation)
capabilities of the estimated models. In Figure 6, the sensitivity and specificity of both
models in relation to WL outcome are plotted for each possible cut-off in the so-called ROC
curves and the Area Under each ROC Curve (AUC) is estimated. AUC measures the
discriminating accuracy of the model, i.e., the ability of the model to correctly classify
patients in their actual quartile of WL.
As a result, the non linear model achieved better results in terms of accuracy and misclassification
rates (70% and 30% vs. 66% and 34%, respectively) than the linear
model.
Fig. 6. ROC curves for nonlinear and linear models
So far, both linear and nonlinear predictive models were built by considering all patients of
the data set, i.e., each model was estimated from a database with known input and output
data.
After this model-building step, the linear and nonlinear models were applied to new
patients, with unknown output values, in order to have a quantitative check on the
effectiveness of the proposed method on the correct selection of the therapeutic effects.
Two additional statistic tools were introduced: the cross-validation method and the
confusion matrix.