Extraction of brainprint by means of autoencoder with attention mechanism

Figure 1. EEG signal preprocessing for channel Fp1. The upper panel shows the raw EEG signal acquired from the Fp1 electrode. The lower panel shows the corresponding preprocessed signal after filtering.

Figure 2. EEG recording session, electrode placement and spatial signal distribution. (Left) Two-dimensional topographic cortical mapping showing the international 10–20 system electrode configuration, with specific localized numerical activation/metric values overlaying each green channel node. (Right) Concurrent multi-channel time-series EEG waveforms mapped from frontal to mastoid channels Fp1 through M2.

Figure 3. Pipeline for converting topomaps into a spatiotemporal EEG data cube. EEG voltages measured at the electrode positions (left) are interpolated to generate a two-dimensional scalp topographic representation (middle, 32 × 32 resolution). Consecutive topographic maps are then stacked along the temporal axis to form a three-dimensional EEG data cube (right, 32 × 32 × 32), preserving both spatial and temporal characteristics of the EEG activity.

Figure 4. Architectural pipeline of the AMAE for feature extraction. This process flow is divided into three sequential phases: encoding, attention and decoding. Initially, an input 3D data cube of size 32 × 32 × 32 is compressed through the Encoder Function network block. The resulting representations are processed via an Attention mechanism to construct a bottleneck feature representation of dimension MxMxM. This optimized feature block is subsequently mapped through the decoder function network to output a reconstructed data cube matching the original input spatial dimensions of 32 × 32 × 32.

Figure 5. Schematic representation of the scaled dot-product attention mechanism. The architectural workflow maps an initial volumetric tensor into a flatten representation. Through parallel linear transformations Q, K and V vectors are acquired. The matrix operation proceeds via a scaled dot production between Q and transpose of K which is normalized using a Softmax. The vector V is weighed by attention weights to generate attention-based feature vector. Finally, this vector is reshaped back into 3D tensor format.

Figure 6. Violin plots for AUC performance of the models across band and stimuli.

Figure 7. Block diagram of autoencoder for resting state EEG records.

Figure 8. Pairwise identification scores for resting state EEG records in Gamma band.

Figure 9. Violin plots for subject identifiability.

Figure 10. Performance evaluation and comparison of 5-subject and 7-subject experiments. (A) SVM based model performance for 5-subject and 7-subject experiment, (B) confusion matrix for 7-subject experiment, (C) confusion matrix for 5-subject experiment.