Bidirectional Reasoning for Channel Selection of Electroencephalography in Auditory Attention Decoding

Abstract

Auditory attention decoding (AAD) from electroencephalography (EEG) is a promising neural sensing approach for neuro-steered hearing devices, but dense electrode configurations limit portability and wearability, making reduced-channel EEG measurement essential for wearable biomedical instrumentation. Current channel-selection methods mainly rely on supervised feature attribution, often yielding model-dependent results that lack cross-perspective consistency and physiological interpretability. To address this, we propose a bidirectional reasoning (BR) framework for reduced-channel EEG configuration, comprising two complementary reasoning streams. The forward stream characterizes intrinsic interchannel relationships using Riemannian-distance analysis, while the backward stream estimates channel contributions to AAD prediction using PlugSelect. The two streams independently generate channel rankings from unsupervised interchannel relationship analysis and label-driven attribution, respectively, and their consistency is used for mutual validation. The results showed that BR identifies similar task-related EEG channels across different perspectives and exhibits cross-modal spatial consistency with functional near-infrared spectroscopy (fNIRS) blood oxygen dynamics, particularly in regions corresponding to the EEG electrode Fp1. These findings provide cross-modal measurement evidence for reliable reduced-channel EEG sensing and support compact electrode configuration design for wearable neuro-steered hearing instrumentation.