Realizing Somatosensory Feedback In A Closed-loop Bi-directional Brain-Computer Interface System For Intelligent Neuroprosthetics Control

Neuroprosthesis is an emerging device that utilizes electrode arrays to interface directly with the nervous systems,having the potential to restore lost motor and sensory functions due to amputation or spinal cord injury(SCI).In the past few decades,research on neuroprosthetics has steadily gained momentum as a result of the increasing consumer base of amputees.However,current neuroprosthetic devices are still very limited in their control ability as well as the lack of sensory feedback which often leads to the absence of a feeling of embodiment by amputees.Additionally,the phantom limb pain(PLP),which is still a major problem with upper limb(UL)amputees is often reported as continuous painful sensations affecting more than 60% of the amputee population.As a result,studies have shown that the rejection rate of UL prostheses are estimated between 35% and 45%.This statistics correctly explains the current situation of the UL prostheses.And if we are to sample the opinion of amputees regarding the high rejection rate,we would realize that the two main research objectives in this field is to improve the motor control,as well as to restore the somatosensory feedback function.To realize this goal,the current dissertation encompasses a closed-loop bi-directional pathways(known as the efferent and afferent),investigating how the braincomputer interface(BCI)technology can be employed to address the main research objectives.I begin by investigating the neural encoding of tactile sensation in small animals using electrical stimulation.By stimulating the peripheral nerves,sensory information are assessed from the surface of the brain cortex with a novel flexible neural interface device for electrocorticography(ECoG)recordings.The effects of varying stimulation parameters such as frequency and intensity were studied,and the outcomes showed a steady and continuous waveforms which establishes a somatosensory evoked potential(SEP)protocol for nerve pathway.Further research was conducted to restore somatosensory feedback by the intervention of sensory event detection.A deep learning model was proposed to detect sensory events captured from brain activity following a mechanical stimulation with Von Frey filaments.Interestingly,the stimulus-evoked sensation achieved better accuracy,sensitivity and specificity.Thus,demonstrating how early detection of sensory events would improve spontaneous sensory and motor activity.In the second part,I proposed an integrated deep learning model consisting of long short term memory(LSTM)and Stacked Autoencoder(SAE)algorithms for decoding motor imagery activities from electroencephalography(EEG)signals acquired from four trans-humeral amputees.The performance evaluation showed a better accuracy,precision,recall,f1?score,specificity and Cohen's kappa.Further analysis with 2-dimensional t-distributed stochastic neighbor embedding(t-SNE)also revealed that the proposed signal decomposition has a distinct multi-class separability in the feature space.Therefore,the study demonstrated the predominance of the proposed model in its ability to accurately classify UL movements from multiple classes of EEG signals,as well as its potential application in the development of an intelligent neuroprosthetic control.Finally,the selection of relevant electrodes that produces optimal subset of EEG and ECoG features is of great importance in BCI applications.Therefore,a neuro-evolutionary algorithm was proposed in the third part of this dissertation to evaluate the channel performance,and subsequently classify the intention tasks of EEG and ECoG signals after feature extraction by common spatial patterns.My findings suggests that optimal selection of relevant BCI electrodes would yield a relatively high accuracy which would facilitate the practical development of a more intuitive and naturalistic prosthetics control.This dissertation has identified and studied some critical issues responsible for high rejection rate and poor performance of the clinically-available UL prostheses such as lack of intuitive and useful feedback.In addition,this dissertation provides potential solutions to those problems as well as valuable insight into the essential requirements for bi-directional BCI,which would facilitate the improvements of neuroprosthetic performance beyond the current application,and also provides possible future research direction.