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

Academic Background

In complex auditory environments, humans can selectively focus on a specific sound source while ignoring other interfering sounds—a phenomenon known as the “cocktail party effect.” Selective Auditory Attention Decoding (AAD) technology decodes the sound source a user is attending to by analyzing brain signals such as Electroencephalography (EEG). This technology has significant applications in neuro-steered hearing aids and Brain-Computer Interfaces (BCIs). However, current AAD algorithms typically rely on supervised learning, which requires users to explicitly indicate their attended sound source to provide “ground-truth labels” for training. In practical applications, obtaining ground-truth labels is often challenging, especially in scenarios where users cannot cooperate (e.g., patients with consciousness disorders). Additionally, the lack of ground-truth labels makes it difficult to evaluate the accuracy of AAD algorithms.

To address these issues, this paper proposes a fully unsupervised method for estimating the accuracy of AAD algorithms. The method is based on the Binary Phase-Shift Keying (BPSK) model from digital communications, treating the AAD decision system as a communication channel with Additive White Gaussian Noise (AWGN), thereby estimating the accuracy of AAD algorithms without requiring ground-truth labels.

Paper Source

This paper was co-authored by Miguel A. Lopez-Gordo, Simon Geirnaert, and Alexander Bertrand. Miguel A. Lopez-Gordo is affiliated with the Department of Signal Theory, Telematics, and Communications at the University of Granada (Spain) and is also a member of the Neuroengineering and Computing Lab (NECOLab) at the Research Centre for Information and Communication Technologies (CITIC-UGR). Simon Geirnaert and Alexander Bertrand are both from the Department of Electrical Engineering (ESAT) at KU Leuven (Belgium), affiliated with the STADIUS Center for Dynamical Systems, Signal Processing, and Data Analytics and the Leuven AI Institute (Leuven.ai). Simon Geirnaert is also part of the EXPLORL research group in the Department of Neurosciences at KU Leuven. This paper was published in the journal IEEE Transactions on Biomedical Engineering in 2025.

Research Workflow

1. Research Objectives and Methods

The objective of this study is to develop a fully unsupervised method for estimating the accuracy of AAD algorithms applicable to correlation-based AAD algorithms in multi-talker scenarios. The method is based on the BPSK communication model, treating the AAD decision system as a channel with AWGN, and estimates the Bit Error Rate (BER) to assess AAD accuracy by estimating relevant parameters such as mean difference and standard deviation.

2. Research Methods and Workflow

2.1 Dataset and Experimental Design

The dataset used in this study includes EEG recordings from 16 normal-hearing participants during a dual-talker task, with a total duration of 72 minutes. The experiment consisted of multiple trials in which participants were instructed to selectively attend to one speaker from the left or right while ignoring the other. The dataset is publicly available online and contains detailed experimental descriptions.

2.2 Speech and EEG Preprocessing

Speech signals were decomposed using a Gammatone filter bank, and subband envelopes were computed using a power-law function to generate the final auditory envelope. Both EEG data and speech envelopes were bandpass-filtered between 1–9 Hz and downsampled to 20 Hz.

2.3 Decoder Setup and Training

The study employed an unsupervised training algorithm to train the stimulus decoder, with initial decoder coefficients set to random values. The integration window of the decoder was set to 0–250 ms post-stimulus, and ten iterations were used to iteratively improve the decoder. Clean speech envelopes were used during both training and testing.

2.4 Unsupervised Accuracy Estimation

The proposed unsupervised accuracy estimation algorithm is based on the BPSK model, calculating BER to estimate AAD accuracy by estimating relevant parameters such as mean difference and standard deviation. The specific steps are as follows: 1. Check the normality of correlation coefficients; if the normality assumption is not satisfied, apply a normalization transformation. 2. Compute relevant parameters (e.g., mean difference and standard deviation). 3. Estimate the mean difference using the folded normal distribution. 4. Calculate BER and accuracy.

3. Experimental Results

3.1 Hypothesis Validation

The algorithm proposed in this paper is based on several assumptions, including normality, uncorrelatedness, equal variance, positive mean difference, and stationarity. Experimental results show that these assumptions hold true in most cases, supporting the effectiveness of the algorithm.

3.2 Results of Unsupervised Accuracy Estimation

Experimental results demonstrate that the proposed unsupervised accuracy estimation algorithm can accurately estimate the accuracy of AAD algorithms. The estimation error remains low across different amounts of training data and decision window lengths. For example, with a 20-second decision window, the mean absolute error is only 3.1 percentage points.

3.3 Application Scenarios

The unsupervised accuracy estimation method proposed in this paper has broad application prospects in neuro-steered hearing aids and BCIs. For instance, in neuro-steered hearing aids, this method can support time-adaptive decoding, dynamic gain control, and neurofeedback. In BCIs, it can provide accurate feedback to caregivers, supporting robust communication paradigms.

Conclusions and Significance

This paper proposes an unsupervised AAD accuracy estimation algorithm based on the BPSK model, capable of accurately estimating the accuracy of AAD algorithms without requiring ground-truth labels. Experimental results show that the method performs well across different amounts of training data and decision window lengths, demonstrating significant scientific value and application potential. This study provides new insights into the unsupervised performance evaluation of AAD systems and lays the foundation for practical applications in neuro-steered hearing aids and BCIs.

Research Highlights

1. Unsupervised Accuracy Estimation: This paper presents the first fully unsupervised method for estimating AAD accuracy, addressing the issue of missing ground-truth labels.
2. Innovation Based on the BPSK Model: By drawing inspiration from the BPSK model in digital communications, this paper treats the AAD decision system as a channel with AWGN, enabling unsupervised accuracy estimation.
3. Broad Application Prospects: The proposed method has significant application value in neuro-steered hearing aids and BCIs, supporting functions such as time-adaptive decoding, dynamic gain control, and neurofeedback.

Other Valuable Information

This study provides important references for future optimization of AAD systems, such as improving decoder performance by increasing the mean difference of correlations. Additionally, the proposed unsupervised accuracy estimation method can be extended to multi-talker scenarios, further expanding its range of applications.