Presentation Open Access
Serious game training versus conventional training in machine learning controlled prosthetic hands: results on functional outcomes Morten B. Kristoffersen1, Andreas W. Franzke1, Alessio Murgia2, Corry K. van der Sluis1, Raoul M. Bongers2
1 University of Groningen, University Medical Center Groningen, Department of Rehabilitation Medicine, Groningen, The Netherlands, 2 University of Groningen, University Medical Center Groningen, Center for Human Movement Sciences, Groningen, The Netherlands,
User training is required to operate machine learning controlled hand prostheses. The goal of training is to make myoelectric activation patterns more distinct from each other and more repeatable. Conventional user training is based on trial and error and is dependent on a phantom hand sensation, which might limit its efficacy. As an alternative, we propose to train using a serious game paradigm that gives feedback on pattern distinctiveness.
The goal of this study is to compare serious game training for machine learning controlled hand prosthesis with conventional training using functional test outcomes.
Participants wore a plaster socket with 8 electrodes. The Michelangelo hand was used and a neural-network regressor calculated the control commands. Participants were assigned to either game or conventional training. Game training consisted of controlling an avatar using a direct mapping from electrode orientation to avatar direction. Each movement that successfully moved the avatar in a certain direction was considered distinct and was used to control the prosthesis. Conventional training followed the principles of Powell (1). Outcome measures were the spherical subset of the SHAP test, the clothespin test and myoelectric pattern distinctiveness (interclass distance). Results are reported using the mean ± SEM without statistical testing.
Six individuals with trans-radial level limb absence participated in 1 fitting session, 2 pre/post-test sessions and 7 training sessions. Preliminary results indicate that for the game training group the SHAP spherical score increased from 20.2 ± 1.3 to 24 ± 5.15 and for the conventional group from 9.8 ± 4.6 to 23.76 ± 3.68 from pre- to post-test respectively. Clothespin time for the game training group decreased from 1:37 ± 0:09 minutes to 1:33 ± 0:04 minutes and for the conventional group it decreased from 2:24 ± 0:16 to 1:11 ± 0:09 minutes. Interclass distance from pre- to post-test increased by 10% ± 7.5% for the game group, while for the conventional group it decreased by 10% ± 12.5%.
DISCUSSION AND CONCLUSION
Preliminary results suggest that both groups improved after training, but that the conventional group improved the most. This might be due to pre-test differences, since the conventional group had very low baseline scores. In the game group interclass distance increased over game playing, showing that the game leads to more distinct patterns. However, the conventional group increased their functional performance while decreasing their interclass distance. This suggests that interclass distance might not fully reflect gains in functional performance.
This work has received funding from the European Unions’s Horizon 2020 programme under grant agreement number 687795 (Project INPUT).
Kristoffersen, 2019, Serious game training in machine learning controlled prosthetic hands.pdf