Adapting Random Forest Classifier Based on Single and Multiple Features for Surface Electromyography Signal Recognition

Surface Electromyography (sEMG) signals have been recently adopted in developing control models for wearable interventional robots and related surgical systems' mechanical control problems. However, adapting sEMG signal models for intravascular robotic interventions is yet to receive any attention due to inability of obtaining continuously reliable signals from the human experts and real-time decoding and integration of the control outputs into the the robotic mechanism. Pattern recognition modeling and classification are vital steps towards developing effective sEMG-based signal models for surgeon-robot intravascular procedures, and these are focused in this paper. For this purpose, nine single features and nine groups of multiple features were investigated for hand motion recognition with the random forest classifier. The proposed random forest classifier was trained and evaluated using a dataset with recordings of twelve finger motions made by ten subjects from a publicly accessible database in the Ninapro project. An average motion classification accuracy of 84.11±3.99% and a best result of subject of 92.94% was obtained from the adapted classifier. Besides this, we also employed three performance metrics to explore more performance detail of the experimental results, and to showed the learning process of the training process.