Explainable and Generalizable Decoding of Auditory Selective Attention from Short, Single-trial EEG for Neurofeedback Training
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Difficulty understanding speech in noisy environments is partly due to problems with auditory selective attention—a process that enhances the neural representation of target sounds over competing sounds. Neurofeedback training that targets auditory selective attention shows promise to improve speech-in-noise perception in normal-hearing individuals. For such training to be effective, the attention decoding algorithm must be accurate, generalizable, and explainable, even for brief, few-second-long EEG signals.Prior attention decoding algorithms used for neurofeedback have low accuracy and lack generalizability, despite their explainability. To address these shortcomings, this study developed three attention decoders based on well-established algorithms: forward linear model, backward linear model, and convolutional neural network (CNN) model. We assessed the decoder’s accuracy in classifying selective attention from single-trial EEG and its generalizability to untrained participants. For explainability, we examined spatial and temporal features critical for decoding.Our findings indicate that decoding selective attention to short, co-located sound is feasible with both linear and CNN models, each offering unique benefits. The CNN showed superior accuracy (72.3%) and cross-subject generalizability compared to linear models (64.9%). Conversely, linear models provided clearer neurophysiological explainability, aligning spatially and temporally with established evidence for attentional modulation of auditory evoked responses.These results support that linear models are more suited for neurofeedback, where understanding the basis of decoding is crucial for effective training. Meanwhile, the CNN remains a strong candidate for applications requiring high performance. This research provides a foundation for developing evidence-based rehabilitation protocols for speech-in-noise perception, enabling interventions grounded in measurable neural changes.