Brain-Computer Interface for Post-Stroke Upper-Limb Rehabilitation: From Simple to Complex Motor Imagery Tasks
Name: CRISTIAN DAVID GUERRERO MENDEZ
Publication date: 11/12/2023
Examining board:
Name |
Role |
|---|---|
| ALBERTO FERREIRA DE SOUZA | Coorientador |
| ANDRES FELIPE RUIZ OLAYA | Examinador Externo |
| DENIS DELISLE RODRÍGUEZ | Examinador Externo |
| TEODIANO FREIRE BASTOS FILHO | Presidente |
Summary: Motor Imagery (MI) of simple tasks such as left and right hand movement, hand opening and closing, and foot or tongue movement has been deeply studied in the literature. Despite the great findings so far, according to the potential use for rehabilitation, there are still many challenges in the scientific community focused on the exploration of more tasks and protocols focused on MI of complex movements, as well as the use of robotic devices for motor assistance considering Activities of Daily Living (ADLs). However, Electroencephalography (EEG)-MI based paradigms have not yet been fully explored in the literature. This master dissertation aims at exploring complex MI tasks assisted mainly by an upper limb exoskeleton and a first-person 2D virtual reality. For this, the perspective from simple MI to complex MI tasks, including those assisted by robotic systems, was evaluated. For simple MI tasks (ST-Set dataset), a public database containing left and right hand MI was used. On the other hand, for exoskeleton MI taks (ET-Set dataset) a proprietary database of 10 healthy subjects was recorded combining MI together with assisted arm flexion and extension movement at two different speeds (30 rpm and 85 rpm). Finally, for MI of complex tasks (CT-Set dataset) a proprietary database of 30 healthy subjects and 7 post-stroke patients was recorded, assisting MI with a first-person 2D virtual reality for the generation of the Action Observation (AO). Different computational techniques were evaluated, including three supervised methods based on Common Spatial Patterns (CSP), two unsupervised method approaches based
on Riemannian Geometry (RG), and three variations of methods based on Deep Learning (DL). Additionally, two classifiers Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM) were evaluated for the supervised and unsupervised methods. Furthermore, two strategies for window segmentation were evaluated. As results, potential performance was found using the
DL methods with Accuracy (ACC) and False Positive Rate (FPR) of approximately 0.6 and 0.45 for ST-Set, 0.98 and 0.05 for ET-Set, and 0.95 and 0.06 for CT-Set (0.8 and 0.22 for post-stroke patients), respectively. Next, RG achieved high performance levels with ACC and FPR of approximately 0.75 and 0.25 for ST-Set, 0.9 and 0.15 for ET-Set, and 0.73 and 0.27 for CT-Set (0.7 and 0.3 for post-stroke), respectively. Finally, the CSP-based methods presented low performance with ACC and FPR of 0.55 and 0.49 for all three datasets. The results allow us to conclude that the presented methodologies of complex MI tasks, as well as the implemented computational variations, are feasible and suitable for the design and implementation of more robust Brain Computer Interface (BCI) systems, allowing a more impactful neurorehabilitation for ADLs recovery in post-stroke patients. In addition, improvements in Human-Machine (HMI) Interaction can be derived, generating increases in restoration due to improvements in usability, controllability and reliability of processes. The novel approaches presented here leave the door open to explore new paradigms, allowing to study the brain effects that occur during these tasks, in order to increase the understanding of the Central Nervous System (CNS).
