Cross-subject seizure detection with isochronous time–frequency gated attention and incremental graph bias

Long-term clinical video-EEG monitoring requires seizure detectors that simultaneously achieve a low false alarm rate (e.g., FA/24h ≈ 1 ), high sensitivity, and low latency under pronounced cross-subject and cross-session shifts. This paper proposes a lightweight event-level framework that keeps all intermediate representations strictly isochronous with physical time and is explicitly optimized for online deployment. The backbone extracts time-domain and frequency-domain features in parallel on a shared time axis, and fuses them by a gated cross-attention module that adaptively balances millisecond transients (spikes, sharp waves) and second-scale rhythm reorganization. On top of this backbone, we build an incremental knowledge graph over subject–channel–band–pattern–window that does not act as a direct classifier, but only generates norm-bounded, temporally smoothed biases for the gates and column-wise attention scores, which can be interpreted as a temporal prior over key times. A small continual-learning stack (replay, distillation, and elastic weight consolidation) and label-free test-time adaptation are further used to mitigate distribution shift while keeping online updates auditable. We evaluate the method on the CHB–MIT dataset under strict online/causal assessment and event-level post-processing, calibrating the decision threshold on validation at a target FA/24h and reusing it on test. Under leave-one-subject-out, the proposed model achieves AUPRC = 0 . 701 , FA/24h ≈ 0 . 98 , Sensitivity = 0 . 861 , and average latency of 5.6 s; in the single-subject setting it reaches AUPRC = 0 . 728 , Sensitivity = 0 . 887 , and latency of 4.9 s. These results indicate that enforcing isochronous time–frequency fusion and using graph-derived temporal priors can improve cross-subject robustness and latency under clinically realistic false-alarm constraints. • A strictly isochronous time–frequency seizure detection framework is proposed. • Gated cross-attention adaptively fuses time-domain and frequency-domain cues. • An incremental knowledge graph provides norm-bounded temporal priors. • The method improves sensitivity and latency under low FA/24h constraints.