Source-linked AI summary
Human brain distinctiveness based on EEG spectral coherence connectivity
Daria La Rocca, Patrizio Campisi, Balazs Vegso, Peter Cserti, Gyorgy Kozmann, Fabio Babiloni, Fabrizio De Vico Fallani
TL;DR
EEG biometric systems often rely on single-region features and may miss dependencies between brain signals. This study fuses spectral-coherence connectivity across brain regions and evaluates it in 108 subjects during eyes-open and eyes-closed resting states. Functional-connectivity fusion achieves higher distinctiveness than power-spectrum measurements, including perfect recognition in frontal regions.
Problem
EEG biometric analyses commonly use single-region power-spectrum features, neglecting temporal dependencies between EEG signals and functional coupling between brain regions.
Method
The study estimates spectral coherence between EEG signals and integrates connectivity information at the match-score level across brain regions.
Results
100% recognition accuracy is obtained for 108 subjects using spectral-coherence fusion in frontal regions during both eyes-closed and eyes-open resting states.
Takeaways & Limitations
Functional-connectivity patterns represent effective features for improving EEG-based biometric systems.
Takeaways & Limitations
Spectral coherence requires EEG signals to be approximately stationary, so short windows or alternative connectivity methods may be needed in other settings.
Abstract
from arXiv · showhide
The use of EEG biometrics, for the purpose of automatic people recognition, has received increasing attention in the recent years. Most of current analysis rely on the extraction of features characterizing the activity of single brain regions, like power-spectrum estimates, thus neglecting possible temporal dependencies between the generated EEG signals. However, important physiological information can be extracted from the way different brain regions are functionally coupled. In this study, we propose a novel approach that fuses spectral coherencebased connectivity between different brain regions as a possibly viable biometric feature. The proposed approach is tested on a large dataset of subjects (N=108) during eyes-closed (EC) and eyes-open (EO) resting state conditions. The obtained recognition performances show that using brain connectivity leads to higher distinctiveness with respect to power-spectrum measurements, in both the experimental conditions. Notably, a 100% recognition accuracy is obtained in EC and EO when integrating functional connectivity between regions in the frontal lobe, while a lower 97.41% is obtained in EC (96.26% in EO) when fusing power spectrum information from centro-parietal regions. Taken together, these results suggest that functional connectivity patterns represent effective features for improving EEG-based biometric systems.
I. INTRODUCTION
EEG biometrics can exploit functional coupling between brain regions, addressing limitations of single-region features and potentially improving recognition across resting-state conditions.
- Resting-state EEG is practical for biometrics because it requires no active subject involvement, reducing inconvenience, fatigue, and artifacts.
- Existing resting-state EEG biometric studies require improved correct recognition performance and larger subject samples.
- Functional connectivity captures temporal dependence between activities in different brain areas, which can provide complementary information to single-electrode features.
- Spectral coherence is less sensitive to EEG amplitude changes and may improve classification under large intra-subject variability.
- The study addresses the tendency to classify from single elements by integrating information from multiple elements through match-score fusion.
- The proposed combined spectral-coherence and classification approach is presented as a first use of this combination for biometric purposes.
A. Dataset and Preprocessing
The study analyzes resting-state EEG from 108 subjects in eyes-open and eyes-closed conditions, segments recordings into short epochs, and extracts power-spectrum features for comparison with connectivity features.
- The dataset contains 108 healthy subjects recorded during 1-minute eyes-open and eyes-closed resting-state conditions.
- A 64-channel EEG system was used, with 56 electrodes retained for subsequent analysis.
- Each subject-condition recording was divided into 6 consecutive nonoverlapping 10-second epochs treated as observations of the same mental state.
- Power spectral density was estimated non-parametrically using Welch’s averaged modified periodogram on 1–40 Hz EEG activity.
- Each electrode produced a 40-element PSD feature vector, yielding 56 electrode-level feature vectors per epoch.
C. Functional connectivity
Functional connectivity is estimated with spectral coherence between electrode pairs across 1–40 Hz, producing frequency-domain features that quantify signal synchrony.
- Spectral coherence quantifies synchrony between two stationary EEG signals at a specific frequency.
- Coherence is computed from the cross-spectrum of two channels normalized by their respective autospectra.
- Coherence values range from 0 for no synchrony to 1 for maximum synchrony at a frequency.
- The method uses 1-second Hanning windows to improve the stationarity of segmented EEG signals.
- Each electrode pair is represented by 40 coherence values spanning 1–40 Hz.
- With 56 electrodes, each epoch contains 1540 electrode-pair coherence feature vectors.
D. Classifier
The classifier identifies subjects by comparing observed EEG feature vectors with Gaussian class distributions using a Mahalanobis-distance rule. Recognition performance is assessed with leave-one-out cross-validation and correct recognition rate.
- Observed feature vectors are assigned to subject identities through discriminant classification under a Gaussian mixture assumption.
- Fisher’s Z transforms COH values, while logarithmic transforms are applied to PSD values before classification.
- A pooled covariance matrix replaces class-specific covariance estimates because each subject contributes only 6 epochs.
- Six leave-one-out runs enroll each subject with 5 epochs and test identification on the remaining epoch.
- The predicted identity minimizes Mahalanobis distance to the enrolled class distributions, using the observed vector, class mean, and pooled covariance matrix.
- Correct recognition rate is computed as the average of the diagonal entries of the N × N misclassification matrix.
E. Match score fusion
The fusion procedure combines complementary PSD or COH evidence from multiple scalp elements at the match-score level. Elements are added selectively within cross-validation, retaining only additions that improve recognition performance.
- Single channels for PSD and channel pairs for COH are fused because different elements identify different subject groups.
- A forward-backward selection procedure orders elements by single-element accuracy and retains additions that improve accuracy.
- Fusion performance is averaged across 6 leave-one-out partitions at every tested subset step.
- Figure 1 organizes COH channel-pair results by connectivity within and between frontal, central, and parieto-occipital zones for EO and EC conditions.
III. RESULTS AND DISCUSSION
The results section compares PSD and COH feature characterizations for subject recognition under eyes-open and eyes-closed resting-state conditions. Both conditions are analyzed separately within the same biometric framework.
- PSD and COH estimates are tested as two characterizations of EEG signals for subject recognition.
- Eyes-open and eyes-closed resting-state outcomes are investigated separately and compared across the performed tests.
A. Single-element classification
Single-element classification maps recognition distinctiveness across individual PSD channels and COH channel pairs in eyes-open and eyes-closed conditions. PSD distinctiveness is strongest centrally during EO and parieto-occipitally during EC, while COH results are represented by adjacency matrices.
- Figure 1 represents COH distinctiveness for every channel pair in adjacency matrices whose axes encode EEG-channel positions.
- Single-channel PSD features reach maximum CRR of 86.91% in the central zone during EO.
- Single-channel PSD features reach maximum CRR of 90.49% in the parieto-occipital zone during EC.
- The maps compare PSD channels and COH channel pairs across EO and EC conditions using spatial scalp and adjacency-matrix layouts.
B. Match-score fusion
Match-score fusion substantially improves recognition across cerebral zones, conditions, and features, with spectral coherence reaching perfect recognition in key settings. The resulting discriminative patterns are mainly short-range, while long-range connectivity adds no performance gain.
- Performance improvements: 100% CRR was achieved with COH fusion in EC across all zones and in EO for the frontal zone.The comparison covers frontal, central, and parietal zones under eyes-open and eyes-closed conditions.
- Performance improvements: Match-score fusion produced dramatic improvements over single-element classification for every zone, condition, and feature.The evaluated features were PSD and COH under EO and EC conditions.
- Connectivity patterns: The optimal COH patterns consisted mainly of short-range channel-pair connections, with EC reaching its maximum CRR more rapidly.The patterns showed hemispheric symmetry, including prominent frontal F7−F8 pairs in both conditions.
- Connectivity patterns: Long-range inter-zone connectivity did not improve recognition performance.The authors relate the short-range advantage partly to volume conduction effects affecting spectral coherence measurements.
- Overall outcome: 100% CRR identified all 108 subjects using regional COH features fused at the match-score level, exceeding reported resting-state EEG results of 98.73% and 97.5%.The comparison baselines used datasets of 45 and 40 subjects, respectively.
C. Limitations and possible solution
The approach has methodological and technical limitations outside the presented protocol, including stationarity requirements, Gaussianity assumptions, sensor-placement demands, and offline computation time. Figure 2 compares CRR for single-element classification and match-score fusion across cerebral zones, with bar colors encoding feature and condition.
- Methodological limitations: Spectral coherence requires EEG signals to be approximately stationary, so short windows or alternative connectivity methods may be needed.The authors mention stationarity testing and wavelet-based methods as possible solutions.
- Figure 2: Figure 2 plots CRR on the y-axis against cerebral zones on the x-axis, with bar colors encoding spectral feature and condition.It compares single-element classification with match-score fusion.
- Technical limitations: The Mahalanobis classifier assumes Gaussian-distributed features, requiring transformations or alternative classifiers when Gaussianity is not met.Reduced polynomial regression-based classifiers are given as one alternative.
- Technical limitations: Match-score fusion requires many scalp sensors, increasing system-design complexity and the time needed to establish skin-sensor contact.Dry miniaturized helmets and non-contact biosensors are proposed as possible solutions.
- Technical limitations: Offline enrollment and fusion-step definition take around 20 minutes on a standard personal computer.The authors suggest parallelizing ranking and cross-validation to reduce computation time.
IV. CONCLUSIONS
The study addresses declining EEG biometric performance at larger subject counts by combining spectral coherence connectivity through match-score fusion. Results support connectivity-based features as effective candidates for robust EEG user recognition.
- EEG biometric classification performance can decrease when the number of people to recognize exceeds 100.
- The fusion procedure combines channel-pair coherence features by retaining steps that improve overall correct recognition rate.Table I organizes correct recognition rates by fusion step, channel pair, condition, and cerebral zone.
- N=108 subjects were evaluated with a high-density EEG system using 56 available electrodes.
- 100% recognition was achieved in both eyes-closed and eyes-open resting states using 15 frontal EEG sensors.The authors identify this reduced-sensor configuration as a technical advantage for future biometric helmets.
- Connectivity-based approaches are proposed as effective invariant features for developing robust EEG-based user-recognition systems.