Bayesian Machine Learning Methods for Brain-Computer Interface Applications
Guardado en:
| Udgivet i: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Hovedforfatter: | |
| Udgivet: |
ProQuest Dissertations & Theses
|
| Fag: | |
| Online adgang: | Citation/Abstract Full Text - PDF |
| Tags: |
Ingen Tags, Vær først til at tagge denne postø!
|
| Resumen: | Biomedical data often pose unique challenges due to their complexity, with the P300-based Brain-Computer Interface (BCI) as an example. The P300 BCI enables communication for individuals with severe physical impairments by interpreting electroencephalogram (EEG) signals produced in response to a series of presented stimuli. This dissertation addresses three major challenges in this field. First, P300 BCI systems typically rely on off-the-shelf classification models that are often black-box in nature and lack interpretability. Moreover, modeling EEG signals in P300 BCI applications presents additional challenges, including limited training data, high-dimensional features, complex spatiotemporal dependencies, and low signal-to-noise ratios. These factors highlight the need for an interpretable model designed specifically for P300 BCIs. Second, there is no systematic approach for determining early stopping criteria or dynamically adapting stimulus presentation patterns to improve system efficiency, which is essential for enabling faster communication in real-world applications. Third, reliable uncertainty quantification methods are needed to control the system's error rate and ensure user satisfaction; however, such methods currently remain underdeveloped. This dissertation introduces novel methodologies leveraging Bayesian modeling, reinforcement learning, and conformal inference to address the three challenges. In Chapter 2, we introduce the Gaussian Latent Channel model with Sparse time-varying effects (GLASS), a Bayesian multinomial regression framework that addresses key challenges in P300 classification. GLASS incorporates a latent channel decomposition to alleviate spatial correlations and a soft-thresholded Gaussian process (STGP) to enhance the signal-to-noise ratio. By identifying key EEG channels and response windows, GLASS improves interpretability while maintaining robust classification performance. We develop a gradient-based variational inference (GBVI) algorithm for fast posterior computation. We demonstrate the effectiveness of GLASS through extensive testing on both simulated and real-world datasets. In Chapter 3, we focus on optimizing the BCI-utility of P300 BCIs through Bayesian model-based reinforcement learning. We construct confidence scores for each character based on previous EEG responses and propose a unified learning framework that explicitly maximizes BCI-utility. An actor-critic algorithm is employed to determine the early stopping criteria, while a Gaussian process-based Bayesian model quantifies changes in confidence scores due to any new stimulus, guiding the selection of the next stimulus for dynamic stimulus presentation. To the best of our knowledge, this work is among the first to directly improve BCI-utility within a reinforcement learning framework. Our approach demonstrates significant improvements in communication efficiency through simulations and analyses of real BCI participant data. In Chapter 4, we focus on developing an uncertainty-aware P300 BCI system using split conformal prediction. While existing uncertainty quantification methods often rely on strong distributional and independence assumptions, split conformal prediction provides finite-sample coverage guarantees without requiring specific distributional forms and only assumes exchangeability of data. Our framework constructs confidence sets at the stimulus, half-sequence, and character levels and incorporates lower-level (stimulus and half-sequence) uncertainty into character-level prediction, effectively utilizing the unique data structure of P300 BCIs. These methods enhance communication efficiency by skipping unconfident characters and enabling early stopping when confidence is high. Through extensive simulations and real data applications on the UMDBI dataset, we demonstrate that the split conformal framework consistently maintains pre-specified accuracy thresholds and improves BCI-utility. Together, these contributions advance the statistical modeling and decision-making strategies for P300 BCIs, addressing key challenges in classification, efficiency, and uncertainty quantification. |
|---|---|
| ISBN: | 9798314875872 |
| Fuente: | ProQuest Dissertations & Theses Global |