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

Abstract

Objective: Selective auditory attention decoding (AAD) algorithms process brain data such as electroencephalography to decode to which of multiple competing sound sources a person attends. Example use cases are neuro-steered hearing aids or communication via brain-computer interfaces (BCI). Recently, it has been shown that it is possible to train such AAD decoders based on stimulus reconstruction in an unsupervised setting, where no ground truth is available regarding which sound source is attended. In many practical scenarios, such ground-truth labels are absent, making it, moreover, difficult to quantify the accuracy of the decoders. In this paper, we aim to develop a completely unsupervised algorithm to estimate the accuracy of correlation-based AAD algorithms during a competing talker listening task. Methods: We use principles of digital communications by modeling the AAD decision system as a binary phase-shift keying channel with additive white gaussian noise. Results: We show that the proposed unsupervised performance estimation technique can accurately determine the AAD accuracy in a transparent-for-the-user way, for different amounts of training and estimation data and decision window lengths. Furthermore, since different applications demand different targeted accuracies, our approach can estimate the minimal amount of training required for any given target accuracy. Conclusion: Our proposed estimation technique accurately predicts the performance of a correlation-based AAD algorithm without access to ground-truth labels. Significance: In neuro-steered hearing aids, the accuracy estimates provided by our approach could support time-adaptive decoding, dynamic gain control, and neurofeedback. In BCIs, it could support a robust communication paradigm with accuracy feedback for caregivers.