Paperdoi:10.64898/2026.11.17.767313

ProlateNet: Sampling-rate transfer of convolutional networks with a weights-only certificate for physiological signals

PreprintDated Nov 19, 2026Added Nov 19, 2026

  • C
  • Z
  • L
  • H
  • X

Authors: Cui, Zhu, Li, Huang, Xu

bioRxiv preprint (posted 2026-11-19; not peer reviewed)

TL;DR

Convolutional networks trained at one sampling rate lose accuracy at others. ProlateNet builds its first-layer kernels from prolate spheroidal wave functions so they suppress aliasing at any rate, and gives a weights-only bound on how much a trained model can degrade before it is deployed. On SEED (EEG), Sleep-EDF-20 and PTB-XL its mean worst drop was at most 0.003, versus 0.18 and 0.35 macro-F1 on SEED for FlexConv and SFI convolution.

Abstract

A central goal of physiological signal analysis is to train a model once and deploy it across recording systems. However, these systems sample EEG and ECG at different rates, and a network trained at one rate loses accuracy at the others. Continuous-kernel networks were built to solve this problem, but published ones still fail at unseen rates, and no existing method can certify a trained model before deployment. In this study, we propose ProlateNet, which builds kernels from prolate spheroidal wave functions to suppress aliasing across sampling rates and certifies a trained model from its weights alone. Across fourteen first-layer designs trained at one rate on three public EEG and ECG datasets, the published continuous-kernel layers FlexConv and sampling-frequency-independent (SFI) convolution show mean worst drops of 0.18 and 0.35 macro-F1 on SEED. ProlateNet is the only design that is stable and accurate on all three datasets with an error bound fixed before training, and its mean worst drop is at most 0.003. Without target-rate recordings, its certificate ranks the continuous-kernel designs on SEED by measured transfer with Spearman ρ = 0.92, against 0.68 for energy above the training Nyquist frequency. Overall, ProlateNet shows that controlling aliasing enables convolutional networks to transfer across sampling rates and provides principled guidance for building physiological signal models that are trained once and deployed across recording systems.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: http://creativecommons.org/licenses/by-nc-nd/4.0/

Results on ChatPPG leaderboards

All leaderboards

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

Tasks & signals

Tasks

Sleep staging

Signals

Keywords

Sampling Rate TransferContinuous KernelProlate SpheroidalSlepianAnti AliasingCertificateMulti RatePtb Xl

Datasets used

How it works

How Method works: Setup → Method → TargetSetup: Signals (eeg, ecg). Method: Method. Target: Sleep staging.SetupSignals: eeg, ecgSignalseeg, ecgMethodMethodTargetSleep stagingSleep staging

Cite this paper

Cite this paper with the fields recorded here. Check it against the publisher before submitting.

ProlateNet: Sampling-rate transfer of convolutional networks with a weights-only certificate for physiological signals. (2026). bioRxiv preprint. https://doi.org/10.64898/2026.11.17.767313
@misc{cuixuprolatenets2026prolatenet,
  title         = {{ProlateNet: Sampling-rate transfer of convolutional networks with a weights-only certificate for physiological signals}},
  year          = {2026},
  howpublished  = {bioRxiv preprint},
  doi           = {10.64898/2026.11.17.767313}
}

Added to the ChatPPG research hub