Identification of finger movements using machine learning techniques applied to forearm high density EMG recordings

Prosthesis available on the market gives the upper-limb amputee the possibility of performing a restricted number of gross hand movements. We think that in using an High Density EMG sensor, we can obtain a greater amount of information about the muscular activity. From this we want to detect ner hand movements such as the nger exions. Data comes from the performance of a speci c motor task by 3 healthy subjects acquired with the HD EMG sensor. After an high pass ltering, we used an ICA algorithm to detect the signal sources. We calculate ve EMG features (the Mean Absolute Value, the Variance, the Willinson Amplitude, the Kurtosis and theWaveform Length) from 100-ms length windows of the sources. Then we train two classi ers (Support Vector Machines and Naive Bayes) to recognize the nger movements. We also implement a real time version of the recognition algorithm. The introduction of the Markov chain can further improve the recognition accuracy. SVM classi er reaches the best accuracy performances in a inter-subjects mean (from 80.1% to 82.28% with Markov chain). Results in a growing training set size shows an increase in the classi cation accuracy. Among the 5 EMG features the best results comes from Mean Absolute Value, Variance and Willinsion Amplitude. The real time algorithm presents both a good speed and a good recognition accuracy. However we observe a big inter-subject variability. In conclusion we can con rm that this approach to the nger movement detection reaches its goal, especially thinking on real time future developments.