Computational Intelligence in Gait and Movement Pattern Analysis - Computational Intelligence in Biomedical Engineering - page 312

Test results indicated that all three NN algorithms provided good classi-
fication accuracy (83%); however, the Bayesian regularization gave superior
classification performance in ROC area, sensitivity, and specificity (Table 7.1).
Performance of the Bayesian regularization was significantly enhanced when
feature selection was applied, that is, a subset of features was selected based on
their relative contribution to the classification tasks. The advantage of feature
selection is illustrated by the increased ROC area as shown in Figure 7.12.

TABLE 7.1
Comparison of Performance of Neural Network Algorithms for the
Classification of Young-Old Gait Using ROC Area and Sensitivity Results

Neural Network Back Scaled Conjugate Bayesian

Algorithms Propagation Gradient Regularization

ROC area 0.82 0.84 0.90

Sensitivity at 0.60 0.67 0.73

specificity of 0.9
and 0.75 0.75 0.79 0.86
Source : Adapted from Begg, R. K. and Kamruzzaman, J., Aus. Phy. Eng. Sci. Med ., 29(2),
188-195, 2006.

BR ROC plots using all and selected features

1

0.9

0.8

All 24 features
3 Best features

0.7

0.6

0.5

0.4 0

0.1 0.2 0.3 0.4 0.5 0.6

1-Specificity

FIGURE 7.12
ROC plots of BR classifier using 24 features and 3 key features selected by the
feature selection algorithm. (Adapted from Begg, R. K. and Kamruzzaman, J.,
Aus. Phy. Eng. Sci. Med., 29(2), 188-195, 2006.)