Unsupervised Accuracy Estimation for Brain-Computer Interfaces Based on Selective Auditory Attention Decoding.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Selective auditory attention decoding (AAD) algorithms use electroencephalography (EEG) to identify attended sound sources among competing sounds.
• Current unsupervised AAD training lacks ground truth, hindering performance quantification.
• Practical applications like neuro-steered hearing aids and brain-computer interfaces (BCI) require accurate decoder performance metrics.

Purpose of the Study:

• Develop a completely unsupervised algorithm to estimate the accuracy of correlation-based AAD algorithms.
• Address the challenge of performance quantification in unsupervised AAD settings.
• Provide a transparent-for-the-user method for AAD accuracy estimation.

Main Methods:

• Model the AAD decision system using principles of digital communications, specifically as a binary phase-shift keying channel with additive white Gaussian noise.
• Employ a correlation-based approach for AAD.
• Validate the unsupervised performance estimation technique across various training data amounts, estimation data lengths, and decision window sizes.

Main Results:

• The proposed unsupervised technique accurately determines AAD accuracy without ground truth labels.
• The method demonstrates transparency and effectiveness across different experimental parameters.
• The approach can estimate the minimal training data required for achieving specific target accuracies.

Conclusions:

• The developed unsupervised estimation technique reliably predicts the performance of correlation-based AAD algorithms.
• This method overcomes the limitation of needing ground truth labels for performance evaluation.
• Accuracy estimates can enhance neuro-steered hearing aids (e.g., adaptive decoding, gain control, neurofeedback) and BCIs (e.g., robust communication, caregiver feedback).