IEEE Robotics & Automation Magazine - December 2017 - 110

Table 2. The classification accuracies obtained in this study for three grips, six grips, eight motion
classes, and ten motion classes.
Groups of Motion
Classes

FSR Placement A, Without
EMG

FSR Placement B, Without
EMG

FSR Placement B, with EMG

Three classes

94.9 ± 1.7%

94.2 ± 2.3%

94.8 ± 2.3%

Six classes

70.7 ± 5.3%

80.4 ± 5.7%

81.6 ± 5.5%

Eight classes

83.0 ± 2.0%

77.9 ± 3.8%

78.5 ± 3.6%

Ten classes

74.1 ± 3.7%

74.3 ± 3.3%

75.2 ± 3.3%

subjects. In this study,
we obtained a maximum
machine-learning accurarecognition, the number of cy of 83 ± 2% for the eight
motion classes (FSR concontrollable motion classes figuration A: 58 FSRs)
and a similar dynamic
protocol using one submay be increased.
ject with transradial amputation. The decrease in
classification accuracy is believed to be due to the weight of
the prosthetic socket and the prosthetic hand and also to a
degeneration of the forearm muscles in the amputee subject, which resulted in the production of less distinguishable signals.

By utilizing pattern

Addition of EMG
With the addition of EMG signals as extra features to FMG
signals, there was a consistently slight increase in classification accuracies for all groups of motion classes. The

Output Class

Confusion Matrix
1

896
33.2%

2
0.1%

0
0.0%

99.8%
0.2%

2
0.1%

656
24.3%

38
1.4%

94.3%
5.7%

2
0.1%

242
9.0%

862
31.9%

77.9%
22.1%

99.6%
0.4%

72.9%
27.1%

95.8%
4.2%

89.4%
10.6%

2
3
Target Class

Figure 10. The confusion matrix for the three-class problem used
for the 2016 Cybathlon. Classes 1, 2, and 3 are relax, open, and
force, respectively.

110

IEEE ROBOTICS & AUTOMATION MAGAZINE

DECEMBER 2017

increase in mean accuracy for the different categories
ranged from 0.6 to 1.2%. Also, the Student's t-test
(p = 0.05) did not prove a significant improvement in
results after the addition of EMG signals. This result shows
that FMG can capture most of the information EMG can
and thus shows high potential as an alternative to EMG. We
recommend a study to compare the following situations: 1)
sensor placement configuration B with EMG (embedding
EMG electrodes in this setting limits the number of active
sensors in the sensor strips placed under EMG electrodes,
which is the setup used in our experiment) and 2) configuration B without EMG cutouts (in this setting, the sensor
strips can be fully extended to the end of the inner socket
toward the elbow). The second was not investigated in this
article because the inner socket used for this study contained cutouts for EMG electrodes.
Cybathlon
The device described in this study was used to participate in
the upper-extremity category of the 2016 Cybathlon. This
category consisted of six main tasks representing ADL.
Based on the information provided by Cybathlon organizers, these tasks were recreated to use for training before the
competition. Then, a task-specific dynamic training protocol was employed. The training consisted of six sets of data,
each including three grips: relax, open, and close. Each of
the grips was held for 15 s as the subject was mirroring the
performance of one of the competition tasks while using our
recreated setup. All six data sets were then used to train the
model used on the day of the competition. The confusion
matrix in Figure 10 shows the classification accuracy
obtained for the three grips that applied the training data
used for the competition. The analysis appropriated the
same LOOCV technique used in all of the offline analysis
reported in this study.
Conclusions and Future Work
Among others, the ease of control and reliability are some
of the reasons for the high rejection rate of powered prostheses [4]. In this study, we introduced potential solutions
by developing a practical prosthetic device that can work
with either EMG, FMG, or both. Increasing control accuracy through the inclusion of EMG signals as extra

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2017