Abstract
Transformer-based analysis of physiological signals depends critically on how input tokens are constructed, yet conventional tokens derived from raw temporal windows or spectrogram patches may overlook the intrinsic oscillatory organization of electroencephalographic (EEG) activity. To address this limitation, we propose AMST-Former, an adaptive mode-wise synchrosqueezing Transformer for patient-independent epileptic seizure detection. Each EEG channel is first decomposed into oscillatory modes using Adaptive Local Iterative Filtering (ALIF), after which the Synchrosqueezing Transform (SST) is applied independently to each mode. This decomposition-before-synchrosqueezing strategy reduces cross-component interference in instantaneous-frequency estimation while preserv- ing mode-specific time–frequency dynamics. The resulting time–frequency patches are embedded together with channel, mode-frequency, and positional information. A hierarchical factorized-attention mechanism then models token relevance across time–frequency regions, oscillatory modes, and EEG channels before a Transformer encoder captures their interactions. Evaluation using a repeated patient-independent five-fold cross-validation protocol on the CHB-MIT database (22 subjects) yielded an F1-score of 70.76% and an area under the precision–recall curve (AUPRC) of 0.870. At the event level, AMST-Former achieved a sensitivity of 81.4% at one false detection per 24 h. Ablation experi- ments showed that mode-wise SST provided more effective representations than direct SST, per-mode Hilbert spectra, and fixed filter-bank decomposition. Furthermore, the factorized-attention mechanism provides mode, time–frequency, and channel-level explanations, whose faithfulness was assessed through deletion analysis. These results support adaptive mode-wise time–frequency tokenization as a promising strategy for interpretable, patient-independent EEG seizure detection.