Work

One-pass restoration and translation of cardiovascular signals with diffusion-trained state-space networks

Year: 2026
Type: preprint
Abstract: Cardiovascular monitoring often relies on recordings degraded by noise or interrupted by signal loss, and some modalities may be unavailable. Existing learning-based approaches can recover missing waveforms, but unified diffusion models can require repeated network evaluations at inference. A bidirectional state-space network trained on diffusion-perturbed observations can instead reconstruct a waveform in one deterministic forward pass. DiBiMa-Cardio and its multimodal extension apply this approach to electrocardiography, photoplethysmography and arterial blood pressure across 57 configurations spanning denoising, imputation, forecasting, super-resolution and cross-modal translation. In a matched training comparison, diffusion perturbation and its step embedding reduced denoising error but increased super-resolution error in most configurations. In a separate inference comparison, one-pass reconstruction yielded lower normalised mean absolute error than the tested iterative sampler in 15 of 18 single-input configurations, with no significant difference in the other three. On 19,996 test windows, the multimodal model reduced normalised error relative to the base model in 53 of 57 configurations under the harder corruption protocol. The multimodal model processed a four-second window in approximately 105 milliseconds at batch size one on the test GPU. Restoring corrupted ECG recovered much of the classification performance lost to corruption, while agreement between translated and recorded waveforms did not consistently preserve predicted diagnostic classes.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7235640921
Cites: 34
Cited by:
Related to: 10