Riemann-Based Algorithms Assessment for Single- and Multiple-Trial P300 Classification in Non-Optimal Environments.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• The P300 wave is a key neural signal utilized in Brain-Computer Interface (BCI) technology for its high bit rate potential.
• Riemannian Geometry-based P300 classification pipelines offer competitive accuracy without spatial filters and require minimal calibration data.

Purpose of the Study:

• To evaluate the effectiveness of Riemannian Geometry-based P300 detection pipelines in non-optimal, noisy environments.
• To compare the performance of five distinct P300 detection pipelines, including three utilizing Riemannian Geometry.

Main Methods:

• Comparison of five P300 detection pipelines: Linear Discriminant Analysis (LDA), Bayesian Linear Discriminant Analysis (BLDA), Riemannian Minimum Distance to Mean (MDM), Riemannian Tangent Space with Logistic Regression (TS-LogR), and Riemannian Tangent Space with Support Vector Machine (TS-SVM).
• Assessment of single-trial and multiple-trial classification accuracies across fifteen subjects.
• Analysis of P300 waveforms and subject-reported focusing methods via questionnaire.

Main Results:

• Bayesian Linear Discriminant Analysis (BLDA) achieved the highest average single-trial accuracy (72.13%), followed closely by Riemannian Tangent Space with Logistic Regression (TS-LogR) (69.22%).
• BLDA demonstrated faster convergence to higher average accuracies in multiple-trial classification, with TS-LogR also showing strong performance.
• Riemannian methods, particularly TS-LogR, proved viable for single-stimulus detection, with accuracies only slightly below BLDA.

Conclusions:

• Bayesian Linear Discriminant Analysis (BLDA) and Riemannian Tangent Space with Logistic Regression (TS-LogR) are identified as viable and high-performing methods for P300 detection in Brain-Computer Interfaces, especially when high bit rates are critical.
• The study validates the utility of Riemannian Geometry-based approaches in challenging, non-optimal BCI environments.
• Further investigation into TS-LogR confirms its potential as a robust classification method for P300 detection.