Computational Intelligence in Gait and Movement Pattern Analysis - Computational Intelligence in Biomedical Engineering - page 317

Polynomial kernel ( C = 1, degree = 3)

2.5

2

1.5

1

0.5

0

− 0.5

− 1

− 1.5

− 2 − 1 0 1 2 3

Feature 13

FIGURE 7.14
Two-dimensional plot showing decision surface for two features, that is, push-
off force and knee range of motion. (From Begg. R. K., Palaniswami, M., and
Owen, B., IEEE Trans. Biomed. Eng. , 52, 828-838, 2005b. c IEEE.)

Figure 7.14 illustrates a 2-D plot of the decision surface for the young and
elderly participants using two gait features from the MTC histograms.
Early detection of locomotor impairments using such supervised pattern-
recognition techniques could be used to identify at-risk individuals. Corrective
measures can then be implemented to reduce the falls risk. Owing to superior
performance of the SVM for gait classification (Begg et al., 2005b) and in
other biomedical applications, SVMs hold much promise for automated gait
diagnosis.
The generalization performance of a classifier depends on a number of fac-
tors but primarily on the selection of good features, utilizing input features
that allow maximal separation between classes. A feature-selection algorithm
was employed by Begg et al. (2005b), which iteratively searched for features
that improved the identification results. The results, from a number of training
and test samples, clearly showed that a small subset of selected features is more
effective in discriminating gait than using all the features (see Figure 7.15).
Classification performance of the SVM depends on the value of the regular-
ization parameter, that must be carefully selected through trial and error.