Aggregation of sparse linear discriminant analyses for event-related potential classification in brain-computer...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Event-related potential (ERP) classification in brain-computer interfaces (BCIs) faces challenges like the curse-of-dimensionality and bias-variance tradeoff.
• Insufficient training samples, common in BCI applications due to limited calibration time, can significantly degrade classification performance.

Purpose of the Study:

• To introduce a novel method, aggregation of sparse linear discriminant analyses (ASLDA), to address the limitations of ERP classification in BCI.
• To enhance classification accuracy and generalization capabilities, particularly in scenarios with limited training data.

Main Methods:

• ASLDA learns multiple sparse discriminant vectors by utilizing l1-regularized least-squares regressions, leveraging the equivalence between LDA and least-squares regression.
• These vectors are aggregated into an ensemble classifier, performing automatic feature selection for dimensionality reduction and variance reduction.

Main Results:

• Extensive comparisons were conducted using three different ERP datasets, evaluating ASLDA against ordinary LDA and other classification algorithms.
• ASLDA demonstrated superior overall performance for single-trial ERP classification, especially when training sample sizes were limited.

Conclusions:

• The proposed ASLDA method is effective in mitigating the curse-of-dimensionality and bias-variance tradeoff in ERP classification.
• ASLDA shows significant promise for improving the practicability of BCIs in small sample size scenarios, enhancing real-world applicability.