Hybrid Correlation-Neural Network Synergy for Gait Signal Classification - Advances in Heuristic Signal Processing and Applications - page 269

Table 12.1 The range or
universe of discourse for the
selected features

Features Range or discourse

[ − 60, 35] cent _ l _ st

msa _ l _ st [8524, 14100]

va _ l _ st [6130, 12878]

cent _ r _ st [ − 59, 35]

msa _ r _ st [8500, 14500]

va _ r _ st [6122, 12900]

cent _ l _ sw [ − 60, 35]

msa _ l _ sw [8442, 14000]

va _ r _ sw [6052,12829]

cent _ r _ sw [ − 60, 35]

msa _ r _ sw [8440, 14040]

va _ r _ sw [6050, 12830]

cent _ ds [ − 59, 35]

msa _ ds [7574, 13920]

va _ ds [5990, 12745]

this process is complete, it is expected that, for a new subject, if the above quan-
tities are calculated and fed to the ERNNs, the system can determine whether the
subject is healthy or not, and also the type of illness, where the subject is found
to be ill. The three features extracted from the left stance interval sequence of a
subject are named as cent _ l _ st , msa _ l _ st , and va _ l _ st . Similarly, the three features
extracted from the right stance interval sequence are named as cent _ r _ st , msa _ r _ st ,
and va _ l _ st , the three features extracted from the left swing interval sequence are
named as cent _ l _ sw , msa _ l _ sw , and va _ l _ sw , and the three features extracted from
the right swing interval sequence are named as cent _ r _ sw , msa _ r _ sw , and va _ r _ sw .
The three features extracted from the double support sequence are named as cent _ ds ,
msa _ ds , and va _ ds . Hence, for each subject under consideration, 15 features are ex-
tracted from five cross-correlograms. Table 12.1 lists these features, used as the
inputs of the ERNN, with their range of values, obtained for the specific problem
under consideration.
The system is configured as a four-class classification system (i.e., C 4) where

the four classes correspond to healthy subjects, pathological subjects suffering from
PD, pathological subjects suffering from HD, and those suffering from ALS. Two
schemes are discussed in this chapter for solving the composite problem, utilizing
time-domain features, where each scheme utilizes more than one ERNN in modu-
lar form [ 13 ]. Each modular ERNN is designed to solve a sub-problem, and these
ERNNs are arranged in a hierarchical fashion where the output of one ERNN de-
termines whether another (or more than one) ERNN should be activated or not.
Each ERNN is activated as a 15-input-1-output system where the 15 inputs are
determined from the features extracted from cross-correlograms, as discussed in
Section 12.4 . For Scheme 1, ERNN1 is trained to solve a binary classification prob-

=