Unsupervised EEG-based decoding of absolute auditory attention with canonical correlation analysis

TL;DR

This work tackles absolute auditory attention decoding (aAAD) from EEG with a fully unsupervised, self-adaptive approach. It builds on discriminative CCA for feature extraction and uses minimally informed LDA (MILDA) with Gaussian mixture thresholding to classify attended versus unattended speech without ground-truth labels, updating the model iteratively as new data arrive. The unsupervised method significantly outperforms state-of-the-art supervised baselines, demonstrates a robust self-leveraging effect, and remains effective under substantial class imbalance and short decision windows, highlighting strong potential for online wearable applications. The results suggest that adaptive, label-free decoding can closely track a listener's attentional state and support real-time auditory attention applications in hearing aids and neurotechnology.

Abstract

We propose a fully unsupervised algorithm that detects from encephalography (EEG) recordings when a subject actively listens to sound, versus when the sound is ignored. This problem is known as absolute auditory attention decoding (aAAD). We propose an unsupervised discriminative CCA model for feature extraction and combine it with an unsupervised classifier called minimally informed linear discriminant analysis (MILDA) for aAAD classification. Remarkably, the proposed unsupervised algorithm performs significantly better than a state-of-the-art supervised model. A key reason is that the unsupervised algorithm can successfully adapt to the non-stationary test data at a low computational cost. This opens the door to the analysis of the auditory attention of a subject using EEG signals with a model that automatically tunes itself to the subject without requiring an arduous supervised training session beforehand.

Figures (6)

• Figure 1: A summary of the unsupervised aAAD algorithm. Blue boxes reflect inputs, yellow boxes contain the modules of the actual algorithm and red boxes the output.
• Figure 2: The average ROC-curves for the unsupervised model proposed in Algorithm \ref{['alg:UnsupCCA']} and the supervised models using normal and discriminative CCA. Remarkably, the unsupervised model outperforms the supervised model. The gray line represents chance level.
• Figure 3: The average auc and standard deviation (shaded area) after training the model with pseudo-labels that are $p_i$ accurate. The model performs significantly better than chance at $p_i = 0.5$. This demonstrates that the model will naturally improve itself at discriminating between attended and unattended EEG. The gray line demonstrates the chance level.
• Figure 4: The ablation study of the proposed unsupervised model shows that the replacement of LDA by MILDA and the introduction of the discriminatory CCA model have the largest impacts on performance. Meanwhile, the transition from supervised to unsupervised CCA only leads to a minimal drop in performance. The blue line is the average AUC over all subjects, each gray line represents a subject.
• Figure 5: As expected, the unsupervised model performs better as the correlation features are estimated on longer window lengths. Nevertheless, it also performs better than chance on small window lengths. The solid line represents the average and the shaded area represents the standard deviation. The gray line is the significance level.