Learning from Adaptive Oscillatory Modes: An Interpretable Synchrosqueezing Transformer for Patient-Independent Seizure Detection

Source abstract

Transformer-based analysis of physiological signals critically depends 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 preserving 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) yields an F1-score of 70.76% and an area under the precision–recall curve (AUPRC) of 0.870. At the event level, AMST-Former achieves a sensitivity of 81.4% at an average false detection rate of 1.39 false detections per 24 h. Ablation experiments show that mode-wise SST provides more effective representations than direct SST, per-mode Hilbert spectra, and fixed filter-bank decomposition. Synthetic experiments confirm the cross-component bias predicted by the analysis and shows that the SST of a mode remains accurate when separation is imperfect, whereas its Hilbert instantaneous frequency does not. Furthermore, the factorized-attention mechanism provides mode, time–frequency, and channel-level explanations, whose faithfulness is assessed through a deletion analysis. These results support adaptive mode-wise time–frequency tokenization as a promising strategy for interpretable, patient-independent EEG seizure detection.