TY - GEN
T1 - Covariate shift adaptation in EMG pattern recognition for prosthetic device control
AU - Vidovic, Marina M.C.
AU - Paredes, Liliana P.
AU - Hwang, Han Jeong
AU - Amsüss, Sebastian
AU - Pahl, Jaspar
AU - Hahne, Janne M.
AU - Graimann, Bernhard
AU - Farina, Dario
AU - Müller, Klaus Robert
N1 - Publisher Copyright:
© 2014 IEEE.
PY - 2014/11/2
Y1 - 2014/11/2
N2 - Ensuring robustness of myocontrol algorithms for prosthetic devices is an important challenge. Robustness needs to be maintained under nonstationarities, e.g. due to electrode shifts after donning and doffing, sweating, additional weight or varying arm positions. Such nonstationary behavior changes the signal distributions - a scenario often referred to as covariate shift. This circumstance causes a significant decrease in classification accuracy in daily life applications. Re-training is possible but it is time consuming since it requires a large number of trials. In this paper, we propose to adapt the EMG classifier by a small calibration set only, which is able to capture the relevant aspects of the nonstationarities, but requires re-training data of only very short duration. We tested this strategy on signals acquired across 5 days in able-bodied individuals. The results showed that an estimator that shrinks the training model parameters towards the calibration set parameters significantly increased the classifier performance across different testing days. Even when using only one trial per class as re-training data for each day, the classification accuracy remained > 92% over five days. These results indicate that the proposed methodology can be a practical means for improving robustness in pattern recognition methods for myocontrol.
AB - Ensuring robustness of myocontrol algorithms for prosthetic devices is an important challenge. Robustness needs to be maintained under nonstationarities, e.g. due to electrode shifts after donning and doffing, sweating, additional weight or varying arm positions. Such nonstationary behavior changes the signal distributions - a scenario often referred to as covariate shift. This circumstance causes a significant decrease in classification accuracy in daily life applications. Re-training is possible but it is time consuming since it requires a large number of trials. In this paper, we propose to adapt the EMG classifier by a small calibration set only, which is able to capture the relevant aspects of the nonstationarities, but requires re-training data of only very short duration. We tested this strategy on signals acquired across 5 days in able-bodied individuals. The results showed that an estimator that shrinks the training model parameters towards the calibration set parameters significantly increased the classifier performance across different testing days. Even when using only one trial per class as re-training data for each day, the classification accuracy remained > 92% over five days. These results indicate that the proposed methodology can be a practical means for improving robustness in pattern recognition methods for myocontrol.
UR - http://www.scopus.com/inward/record.url?scp=84929471501&partnerID=8YFLogxK
U2 - 10.1109/EMBC.2014.6944592
DO - 10.1109/EMBC.2014.6944592
M3 - Conference contribution
C2 - 25570960
AN - SCOPUS:84929471501
T3 - 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2014
SP - 4370
EP - 4373
BT - 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2014
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2014
Y2 - 26 August 2014 through 30 August 2014
ER -