PhysioWave-T: Token-Efficient Wavelet Foundation Models for Biosignals

PhysioWave-T. Token-Efficient Wavelet Foundation Models for Biosignals. By Yanlong Chen and four others, from Nanyang Technological University, the University of Modena and Reggio Emilia, and two other institutions. Published on the sixth of November.

Biosignals come in several modalities: EEG, ECG and sEMG. Wavelet decomposition splits a signal into subbands, one more subband than the number of levels used, and if each subband is tokenized on its own, sequence length multiplies by that count across every channel and every temporal patch. Add the fact that recordings differ in their montages and sampling rates, and a single foundation model has both a length problem and a setup problem to solve.

PhysioWave-T is a family of token-efficient wavelet foundation models aimed at both. Its central move is dynamic scale folding, which recombines the wavelet subbands before tokenization. Sequence length then scales with channels times temporal patches alone, and the multiplier from the levels disappears.

For the setup problem, route-specific front ends and learned embeddings of channel names let one Transformer per modality handle diverse montages and sampling rates. A setup the model has never seen reuses the backbone through lightweight adapters. The authors pre-train a thirty million parameter EEG model on over fifteen million windows, with anywhere from nineteen to one hundred and twenty-eight channels. Alongside it sit an ECG encoder of thirty million parameters and an sEMG encoder of twelve million.

A further piece, reliability-aware latent fusion or RALF, combines the summary tokens of encoders that were pre-trained independently, using weights that adapt to each sample.

Across four EEG routes, scale folding cuts Transformer latency by between four and seven point two times compared with scale expansion. Peak memory drops by between six point eight and twelve point seven times. Neither comes with a loss of accuracy. Against BENDR, BIOT, LaBraM and EEGPT, the EEG encoder leads on eight of nine metrics, and it surpasses LUNA-L on wearable seizure detection while using fewer parameters.

On SeizeIT2, RALF raises event sensitivity at eleven false alarms per hour from sixty-one point four percent to seventy-four point four percent over the best unimodal baseline. Window AUPRC rises from zero point three two one to zero point three nine nine. It beats weighted-logit and concatenation fusion in every variant. The family also reports state-of-the-art sEMG accuracies, and the top macro-F1 on PTB-XL Super and on Chapman-Shaoxing.

Everything here hangs on where the folding happens. Doing it before tokenization is what turns a sequence inflated by its wavelet levels into one sized by channels and time, and the latency and memory gains follow from that single choice, with accuracy held where it was.