Decoding Absolute Auditory Attention from EEG Signals Using Parameterized Power Spectra

Detecting whether a listener is attentive to auditory input, a process known as absolute auditory attention decoding (aAAD), is a critical prerequisite for selective attention applications but remains challenging, especially in short EEG segments. Existing features, such as spectral entropy (SE) or neural envelope tracking, often provide limited accuracy. We introduce a five-parameter power spectral density (PSD) representation comprising aperiodic (Exponent and Offset) and periodic (Centre Frequency, Bandwidth, and Peak Power) components extracted using the FOOOF algorithm. We recorded EEG data from 40 participants across four auditory tasks: one active and three disengaged (distraction, distancing, and inhibition). We grouped disengagement trials as the negative class. On 6-second EEG windows, Random Forest classification using aperiodic features alone achieved 86.3 % accuracy, significantly outperforming the SE baseline (71.2%). Combining aperiodic and periodic features further boosted accuracy to 95.1 %, cutting the window length typically required for SE in half while exceeding its performance. We evaluated both subject-pooled and subject-dependent fivefold cross-validation schemes, with subject-dependent accuracy reaching 97.4 %, confirming robust performance across both validation frameworks. These findings demonstrate that aperiodic 1/f features offer more robust and interpretable markers of auditory attention than traditional oscillatory measures, enabling efficient real-time attention monitoring in cognitive and clinical applications.