One wavelet backbone for EEG, ECG and EMG: folding subbands cuts the sequence fourfold and 128-channel memory from about 32 GB to 2.5 GB

Biosignal hardware varies widely: EEG runs from two-channel wearables to 128-channel scalp arrays, ECG from single-lead patches to twelve-lead systems, and sEMG from wristbands to high-density grids, each with its own sampling rate and electrode geometry; PhysioWave-T lets one Transformer backbone per modality take these configurations directly. Previously each PhysioWave encoder was confined to a single fixed input configuration, and its tokenizer expanded every wavelet scale into a separate token row, so channel count, window duration and decomposition depth lengthened the sequence together. The pre-trained EEG encoder has 30M parameters and covers more than 15 million four-second windows at 19 to 128 channels and 256 or 512 Hz; across sleep staging, P300 and error-related negativity detection, the fine-tuned encoder reaches the highest value on eight of nine metrics.

Multiscale information no longer costs sequence length: dynamic scale folding recombines the wavelet subbands, with data-dependent weights, into one signal shaped like the input before tokenization, so deeper decomposition only enriches each token's content. Scale expansion sends every subband to the backbone at four times the folded sequence length; a static uniform fold scores 1.6 points lower on average and 1.3 points below scale expansion. Across four EEG routes, folding reduces Transformer latency by 4.0–7.2 times, peak memory by 6.8–12.7 times and FLOPs by 4.3–5.5 times; on the 128-channel route a batch drops from about 32 GB to 2.5 GB, and the fold itself accounts for only 1.5–2.6% of the forward pass.

Encoders pre-trained independently per modality can be fused with per-sample reliability for wearable seizure monitoring, without joint multimodal pre-training. PhysioWave combined unimodal logits with one learned weight per modality, fixed after training, so it could not reflect a particular sensor degrading in a particular window. On SeizeIT2, which records behind-the-ear EEG, ECG and EMG simultaneously, fusion raises event sensitivity at 11 false alarms per hour from 61.4% for fine-tuned EEG alone to 74.4%; window-level AUPRC rises from 0.321 for the best unimodal model to 0.399, and all six training variants beat weighted-logit and concatenation fusion whenever EEG and ECG are jointly present.

The demonstrated scope is three signal types: EEG covering pre-training routes of 19 to 128 channels at 256 and 512 Hz, with downstream sleep staging, P300, error-related negativity detection and two-channel behind-the-ear seizure detection; ECG at twelve leads and 500 Hz; and sEMG pre-trained at 16 to 64 channels and 2 kHz and transferred to an eight-electrode 200 Hz wristband and a 28-electrode 2 kHz array. The multimodal fusion experiments use frozen encoders connected to wearable acquisition configurations through task-specific front ends and adapters.

The sensor corruptions examined in multimodal fusion are injected noise, baseline drift, saturation, bad channels and temporal dropout rather than natural failures in real recordings; the authors list prospective recordings, joint multimodal pre-training and new routes such as photoplethysmography as next steps. On the error-related negativity task the frozen EEG encoder falls below a model trained from scratch, and fine-tuning the backbone is needed to surpass it.