Recursive Estimation of User Intent from Noninvasive Electroencephalography using Discriminative Models

Niklas Smedemark-Margulies, Basak Celik, Tales Imbiriba, Aziz Kocanaogullari, Deniz Erdogmus

Abstract

We study the problem of inferring user intent from noninvasive electroencephalography (EEG) to restore communication for people with severe speech and physical impairments (SSPI). The focus of this work is improving the estimation of posterior symbol probabilities in a typing task. At each iteration of the typing procedure, a subset of symbols is chosen for the next query based on the current probability estimate. Evidence about the user's response is collected from event-related potentials (ERP) in order to update symbol probabilities, until one symbol exceeds a predefined confidence threshold. We provide a graphical model describing this task, and derive a recursive Bayesian update rule based on a discriminative probability over label vectors for each query, which we approximate using a neural network classifier. We evaluate the proposed method in a simulated typing task and show that it outperforms previous approaches based on generative modeling.

Introduction

For individuals with severe speech and physical impairments (SSPI), noninvasive Brain-Computer Interfaces (BCIs) utilizing Electroencephalography (EEG) represent a critical communication lifeline. A prevalent paradigm is the Rapid Serial Visual Presentation (RSVP) speller, where users observe a rapid sequence of stimuli, and the system detects specific brain responses—such as the P300 wave—associated with target symbols.

Traditionally, these systems rely on generative models that attempt to characterize the distribution of the EEG signal given a target ($p(\text{signal} | \text{target})$). However, EEG data is high-dimensional, noisy, and exhibits significant inter-subject variability. Generative approaches often require aggressive dimensionality reduction, which can inadvertently discard relevant signal components.

Methods: Discriminative Recursive Bayesian Updates

This work proposes a shift toward discriminative modeling for updating beliefs about user intent. Instead of modeling the signal distribution, a discriminative model directly estimates the probability of a target given the observed EEG ($p(\text{target} | \text{signal})$).

Graphical model of the recursive Bayesian update process for user intent estimation.

The core technical contribution is the derivation of a recursive Bayesian update rule that integrates standard discriminative classifiers, such as Neural Networks, into a Bayesian framework. This approach allows the system to process high-dimensional, raw EEG data without heavy preprocessing, update symbol probabilities after every stimulus, and utilize modern deep learning architectures like EEGNet and Convolutional Neural Networks (CNNs).

Datasets and Experimental Setup

The proposed method was evaluated using a large-scale RSVP benchmark dataset containing over one million stimulus presentations. This substantial dataset enables the effective training of high-capacity deep neural networks, which is often a bottleneck in BCI research due to the "messy" nature of EEG data.

Key Results

The discriminative approach was compared against traditional generative baselines, such as Linear Discriminant Analysis (LDA) with Kernel Density Estimation.

Comparison of discriminative and generative models across balanced accuracy and Information Transfer Rate (ITR).

As shown in the results, discriminative models significantly outperformed generative ones:

  • Accuracy: The 1D CNN achieved a balanced accuracy of 78.2%, compared to 68.7% for the best generative baseline.
  • Information Transfer Rate (ITR): The 2D CNN reached an ITR of 1.153 bits per symbol, nearly doubling the performance of the generative LDA baseline (0.678 bits/symbol).
  • Scaling: A clear trend emerged where increasing model capacity (parameter count) generally improved communication speed, suggesting that BCI performance is currently limited by model power rather than just signal quality.

Discussion and Future Work

While the results are promising, several avenues for improvement remain. This study utilized uniform priors; integrating language models to account for symbol frequencies and sequences (e.g., 'u' following 'q') could further enhance performance. Additionally, while the simulated typing task provides strong validation, future work must demonstrate these gains in real-time, closed-loop sessions. Finally, exploring transfer learning could help adapt these powerful models to new users with minimal calibration data, addressing the challenge of inter-subject variability in EEG signals.

Summary updated

Abstract

We study the problem of inferring user intent from noninvasive electroencephalography (EEG) to restore communication for people with severe speech and physical impairments (SSPI). The focus of this work is improving the estimation of posterior symbol probabilities in a typing task. At each iteration of the typing procedure, a subset of symbols is chosen for the next query based on the current probability estimate. Evidence about the user's response is collected from event-related potentials (ERP) in order to update symbol probabilities, until one symbol exceeds a predefined confidence threshold. We provide a graphical model describing this task, and derive a recursive Bayesian update rule based on a discriminative probability over label vectors for each query, which we approximate using a neural network classifier. We evaluate the proposed method in a simulated typing task and show that it outperforms previous approaches based on generative modeling.
BibTeX
@inproceedings{smedemarkmargulies2023recursive,
  author = {Niklas Smedemark-Margulies and Basak Celik and Tales Imbiriba and Aziz Kocanaogullari and Deniz Erdogmus},
  title = {{Recursive Estimation of User Intent from Noninvasive Electroencephalography using Discriminative Models}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  doi = {10.1109/icassp49357.2023.10095715},
}
Figure