A self-supervised reconstruction framework for segment-level detection of task-induced deviations in multi-channel EEG

Understanding brain activations during mental arithmetic tasks provides valuable insights into cognitive processing. However, analyses of the same electroencephalogram (EEG) dataset have produced inconsistent findings about the cortical regions involved, raising methodological questions about global signal treatment. Detecting task-related activations without pre-labelled datasets further remains challenging when supervised approaches are not applicable. This study presents a self-supervised framework for segment-based anomaly detection using deep learning to identify EEG signal outliers during mental arithmetic. Using a hierarchical variational auto-encoder, the study reconstructed multichannel EEG data segmented into 1-second intervals and trained exclusively on resting-state recordings. The study assessed reconstruction quality using multiple error metrics, including dynamic time warping, short-time Fourier transform loss, and signal-to-noise ratio, applied individually to channel segments. Outlier segments were detected based on the elevated reconstruction error relative to the resting baseline, using an unsupervised detection method. By analysing deviations at the segment level rather than treating the signal globally, the proposed framework captured transient cortical dynamics that may otherwise be obscured. The results demonstrated that these outlier segments exhibited distinct temporal and spectral characteristics, particularly within the delta, theta, and beta bands, and were most frequently located in peripheral electrodes such as Fp2, O1, O2, P3, and T6. Additionally, co-occurrence analysis uncovered functional patterns of simultaneous anomalies across specific electrode pairs and triplets, suggesting temporally dynamic and spatially distributed cognitive processes during arithmetic reasoning. Overall, the proposed framework provided a fine-grained temporal and spatial characterisation of EEG deviations during mental arithmetic and offers a methodological perspective that helped explain discrepancies in previously reported electrode-level findings.