Biomedical Engineering Reference
In-Depth Information
7.4.1.1
Neural Networks in the Recognition
of Simulated Gait
Barton and Lees (1997) applied NNs to identify three gait conditions: normal
walking, leg length difference, and leg weight difference from joint kinematics
data. They used hip and knee joint angles over the gait cycle along with trans-
formed Fourier coecients as inputs to the classifier, and a small sample of five
male and three female subjects was used. The kinematic data were collected
using reflective markers at the neck, hip, knee, and ankle joints, which allowed
the determination of hip and knee angles. The knee and hip angle data were
then normalized to 128 time intervals and the corresponding FFT coecients
were determined. The first 15 Fourier coecients (8 real and 7 imaginary)
were used for NN training and testing to evaluate NN's performance in recog-
nizing the three simulated gaits. A BP-learning algorithm was used to train
the network. The input layer had 30 nodes for accepting FFT coecients
concerning hip and knee angle information, the two hidden layers consisted
of nine nodes, and the output layer had three output nodes that represented
the three simulated gait types. The training data set included six subjects,
whereas the test set had information on two subjects. The correct recognition
ratio of the NN was reported to be 83.3%. Although a small sample was used
and accuracy was moderate, the results suggest that a trained NN is power-
ful enough to recognize changes in simulated gait kinematics; joint angles in
this case.
7.4.1.2
Neural Networks in the Recognition
of Pathological Gait
Several studies have reported the application of NNs to recognizing healthy
and pathological gait patterns, one of the first was by Bekey et al. (1977),
who showed that an EMG-based NN model could distinguish normal from
pathological gaits. Holzreiter and Kohle (1993) applied NNs to differentiate
healthy and pathological gaits using vertical ground-reaction forces recorded
from two force plates; this experiment employed 131 patients and 94 healthy
individuals. The vertical force-time information was divided into 128 constant
time intervals and all the force components were normalized to the subject's
body weight. FFT coecients of the resulting data were calculated, of which
the 30 low (real and imaginary) coecients were used as inputs to the clas-
sifiers. The NN architecture included a three-layer feedforward topology with
30 input nodes and a single-classification output node corresponding to the
healthy or pathological class. The network was trained 80 times using a BP
algorithm with random weight initialization. A different proportion of the data
set (20-80%) was used for training and testing the network performance. The
maximum success rate for the training set that involved 80% of the data was
reported to be
95%. Wu and Su (2000) also used models based on NNs and
foot-ground reaction forces to successfully classify gait patterns of patients
with ankle arthrodesis and healthy individuals.
∼
Search WWH ::
Custom Search