PhysioWave-T: Token-Efficient Wavelet Foundation Models for Biosignals
The authors report 4.0–7.2× lower Transformer latency and 6.8–12.7× lower peak memory on four EEG routes, without accuracy loss.
PhysioWave-T folds wavelet subbands together before tokenization, reducing the sequence length compared with scale expansion. The family includes separate EEG, ECG and sEMG encoders, with lightweight adapters for unseen recording setups. Its EEG encoder leads eight of nine metrics against BENDR, BIOT, LaBraM and EEGPT. On SeizeIT2, reliability-aware fusion raises seizure-event sensitivity from 61.4% to 74.4% at 11 false alarms per hour compared with the best unimodal baseline.
HF Daily Papers' note
PhysioWave-T folds wavelet subbands together before tokenization, reducing the sequence length compared with scale expansion. The family includes separate EEG, ECG and sEMG encoders, with lightweight adapters for unseen recording setups. Its EEG encoder leads eight of nine metrics against BENDR, BIOT, LaBraM and EEGPT. On SeizeIT2, reliability-aware fusion raises seizure-event sensitivity from 61.4% to 74.4% at 11 false alarms per hour compared with the best unimodal baseline.
HF Daily Papers' note
score 5