A Systematic Evaluation of EEG Frequency-Band Combinations for Biometric Subject Identification
Year: 2026
Type: article
Abstract: Electroencephalography (EEG) has been investigated as a biometric modality because subject-specific characteristics embedded in brain activity can support person identification. However, the contribution of different frequency bands and their combinations to biometric performance remains dependent on the experimental condition and evaluation protocol. This study presents a systematic evaluation of canonical EEG frequency-band combinations for biometric subject identification using the Auditory Evoked Potential EEG-Biometric Dataset. The analysis included 120 two-minute recordings from 20 subjects, restricted to two resting-state conditions: eyes open (EX01) and eyes closed (EX02), with three sessions per condition. EEG signals from four channels (T7, F8, Cz, and P4) were represented using relative power in the Delta (0.5–4 Hz), Theta (4–8 Hz), Alpha (8–13 Hz), and Beta (13–30 Hz) bands, resulting in 16 features per recording. All 15 non-empty frequency-band combinations were evaluated using a 300-tree Random Forest classifier and session-aware leave-one-session-out validation, with feature standardization performed using training data only within each fold. Performance was assessed using accuracy and macro-F1, together with bootstrap confidence intervals and paired permutation tests with Holm correction. For EX01, the highest observed accuracy was obtained with All_Bands at 71.67%. For EX02, Theta+Alpha+Beta achieved the highest observed accuracy of 95.00%, compared with 93.33% for All_Bands. This 1.67-percentage-point difference was not statistically significant. A data-integrity audit identified an exact duplicate relationship for one subject in EX02; sensitivity analysis showed that this did not change the principal conclusion. The findings demonstrate that frequency-band selection can substantially affect EEG biometric performance and that multi-band representations can provide strong cross-session identification performance. However, no single frequency-band configuration can be considered universally optimal based on the present dataset and experimental conditions. The proposed controlled evaluation provides a reproducible framework for assessing spectral representations in EEG biometric identification.
Authors: Toumi Assia, Adjoudj Réda
Institution: Université Djilali de Sidi Bel Abbès
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223385964