چكيده به لاتين
One of the major applications of neural signals decoding in nerve prosthetics is to retrieve sensory-motor functions lost in patients who have been spinal injured. The decoding of motor signals is used to make the prosthesis move and the sensory signals are used to create sensory feedback. Electrode implanting and recording of peripheral nerves are more specific to the central nervous system and have less harmful and invasive levels of electrode implanting in the central nervous system. In previous studies in the field of decoding sensory signals, the possibility of classification of two or three classes of sensory stimuli has been studied and the classification of more types of different sensory stimuli has not been studied. Our goal in this thesis is to use different methods to extract the electroneurographic signal features for decoding sensory stimuli, evaluate and compare these methods with each other and provide a classification model with acceptable decoding accuracy.
Surgery and recording of electroneugraphic signal had done with three-channel electrode located around the sciatic nerve for five vistar rats. Four types of sensory stimulation including foot pinching, foot squeezing, fingers and ankle flexion were applied to the hind paw of each rat in different sessions. Temporal features include mean absolute value (MAV), variance (VAR) and wave length (WL) and spectral features includingautoregressive (AR) and Fast Fourier Transform (FFT) coefficients were extracted. In this study, two types of classifier, SVM and KNN classifiers were used for separation of five classes including four aforementioned stimuli and a non-stimulation resting mode, and the performance of the methods was compared.Compared to the two types of classifiers examined, the SVM classifier had better outcomes for all three features than KNN. Compared to the features, the FFT (89.8%), AR (82.38%) had better performance than the time (59.23%). Between two electrodes with gold and copper outer layers, gold has better results. In different channels selection, three-channel averaging has better results.
We showed that, for a minimum 100 milliseconds of signal with spectral features, multiclass classifications could have acceptable accuracy to sensory stimuli.
