Computational Intelligence in Gait and Movement Pattern Analysis - Computational Intelligence in Biomedical Engineering - page 320

basic gait parameters makes it an attractive alternative for identifying CP
children and evaluating the effectiveness of treatment.
7.5 Concluding Remarks
In this chapter, we reviewed the gait pattern recognition and classification
abilities of CI techniques. Gait features used in training and testing the
machine classifiers are usually extracted from data recorded using standard
techniques. Gait features employed in studies to date include kinematic,
kinetic, and EMG variables. Past research supports the use of NNs, fuzzy
logic, GAs, SVMs, or a combination of these (e.g., neurofuzzy) for the recog-
nition of gait types (healthy, ageing, pathological, or faller). When differen-
tiating gaits, performance can be expressed using cross-validation accuracy
rates and ROC plots (ROC area, sensitivity and specificity measures). Clas-
sification performance has been found to be significantly enhanced when a
subset of selected features are used to train the classifiers. Feature selection
appears, therefore, to be important when dealing with gait data classification;
this can also reduce the dimensionality of input data with less computational
complexity.
CI offers benefits for gait data modelling and gait pattern prediction. In
most applications, the reported classification accuracy is reasonably high, sug-
gesting excellent prediction abilities and suitability for modelling the under-
lying gait data structures. Furthermore, performance using CI techniques, for
example, NNs and SVM, is superior to an LDA-based classifier. In studies of
both healthy and pathological populations (Wu et al., 1998; Kamruzzaman
and Begg, 2006), improved performance using CI tools also suggested the
importance of nonlinear approaches for modelling the relationships between
disease or ageing and gait characteristics.
Based on this review, it is reasonable to conclude that further advances
in movement pattern recognition may be achieved through the incorporation
of two approaches. The first is feature selection , including additional feature-
selection algorithms, such as backward elimination (Chan et al., 2002). GAs
appear to have much potential (Su and Wu, 2000; Yom-Tov and Inbar, 2002)
for separating relevant features to improve gait detection and classification.
The second approach which has considerable potential is hybridization , that
is, incorporating more than one artificial intelligence technique (Lauer et al.,
2005) into the classification process. Some examples are combining an NN
with fuzzy logic (neurofuzzy scheme) or evolutionary algorithms (Yao and
Liu, 2004).
Good progress has been made over the past decade in the application of CI
to recognizing gait movement abnormalities, and also establishing CI as a valu-
able diagnostic tool in biomedical science. Further research is needed before
these techniques become routine diagnostic tools within clinical settings.