Electroencephalographic (EEG) biometric features have attracted considerable interest, but they have also drawn major criticism as having low stability. Moreover, most published studies ignore the potential effect of individual factors interacting with stability on the performance of the examined system. We examined the effects of age, sex, and neurological conditions in 60 subjects: 1) using a single EEG recording; 2) when pooling two EEG sessions; and 3) when performing cross-session cross-validation in a biometric system based on multivariate autoregressive measures, in order to extend previous work on autoregressive coefficients. Feature-level fusion and wrapped feature subset-selection in multivariate autoregressive coefficients (MVAR), power spectral density, and 14 additional measures derived from the MVAR model resulted in a maximum area under the curve of receiver operating characteristics (AUC of ROC)/top equal error rate of 98.85/5.34 for a single EEG, 95.51/9.70 for EEG-pooling, and 85.34/22 for cross-session cross-validation. This best result was obtained by the transfer function polynomial, which outperformed the MVAR coefficients, based on the example data set used in this paper. Age, sex, and pathology significantly interacted with the stability of features (p < .001). We suggest further investigation of frequency-dependent measures derived from the MVAR model. We emphasize the serious problem of ignoring stability in most previously published research and recommend accurate reporting of individual factors when studying EEG biometric features in multiple sessions for enrolment and authentication on separate days.
Find related publications in this database (Keywords)EEG biometric features