Random forests based sub-vocal electromyogram signal acquisition and classification for rehabilitative applications

The proposed research focuses on designing a low-cost electromyogram (EMG) data acquisition system (DAQ). The developed system acquires EMG signals from the sub-vocal region and suitable features are extracted using time-frequency transform such as Wavelet Transform. Once the features are extracted, the final classification is carried out using ensemble decision trees called Random Forests (RF). Giving the randomness in the ensemble of decision trees (DT) stacked inside the RF model, this technique can provide at the recall stage, not only the early assessment of classification, but also a probability outcome which quantifies the confidence level of the decision. The performance accuracy is found to be more than 90% when two features were considered compared to 75% with five features. Thus there is a trade-off between the input features versus the classification accuracy. Thus, the proposed data mining based technique will be highly suitable for developing EMG signal acquisition system used for bio-medical instrumentation.