Improving P300 BCI Speller Using Uncertainty Quantification

Renardi, Bernard (2024) Improving P300 BCI Speller Using Uncertainty Quantification. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2024_RenardiB.pdf Download (9MB) | Preview
	Text toestemming renardi.pdf Restricted to Registered users only Download (266kB)

Abstract

Brain-computer interfaces (BCI) allow individuals with impaired motor movements to gain control of communication. One of the applications is the P300 speller, which uses brain signals obtained through electroencephalography (EEG) to control typing. However, challenges exist: low mobility due to wired devices, slow processing speed, and low performance. To address these challenges, we first investigate the feasibility of capturing P300 using dry electrode EEG. Second, we propose an adaptive threshold method to measure confidence levels using UQ. Our model was tested and evaluated using a data set from patients with amyotrophic lateral sclerosis (ALS). Overall, the integration of UQ into our chosen classifier using a Riemannian distance metric significantly outperformed other methods. We demonstrate how UQ can reduce the average number of flashes required and has been shown to improve the overall F1-score across all subjects. However, the performance of dry electrode EEG remains inferior to that of wet electrode EEG. We conclude that addressing the feasibility of using dry electrodes for the P300 remains a challenge that requires alternative approaches to bridge the gap with wet electrodes before achieving daily usability. Additionally, the integration of uncertainty quantification into the P300 speller demonstrates its potential to reduce time requirements and enhance robustness.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Sburlea, A.I. and Jong, I.P. de
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	14 Mar 2024 08:29
Last Modified:	14 May 2025 14:29
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/32009

Actions (login required)

View Item