Machine Learning-based Method to Label Signals from People with Neurological Injuries

Interfaces that use sEMG signals face the challenge of correctly identifyingthe signal while distinguishing it from noise or interference.Although classical techniques like visual inspection and machinelearning methods exist, most studies focus on signals from healthyindividuals. There is a lack of data and methods suitable for signalsfrom individuals with neurological conditions, such as cerebralpalsy and post-stroke. This study analyzes sEMG data from individualswith neurological injuries, using machine learning methods toidentify muscle contractions and rest without pre-processing. Thedata were acquired from people with neurological diseases, such ascerebral palsy and post-stroke. They were extracted using sEMGfrom triceps brachii and extensor carpi radialis muscles. The signalswere not preprocessed and were input as segmented time windowsto three proposed classifiers: Support Vector Machine, Random Forestand an Ensemble Voting classifier. All three classifiers reachedaround 99% accuracy and F1-Score on typical sEMG data, but theresults on abnormal data were inconclusive.