City Research Online

Abstract

Deep learning models for reconstructing arterial blood pressure (ABP) from photoplethysmography (PPG) achieve strong predictive performance but offer limited insight into how this relationship is structured. In this work, we introduce discretisation as a framework for studying intrinsic structural interpretability in the PPG-to-ABP mapping using a VQ-VAE model, and assess its feasibility by testing whether the resulting representations exhibit non-trivial structural organisation. We encode signals as sequences of discrete tokens and evaluate their organisation through four properties: component reuse, consistency of local waveform patterns, sequential dependencies, and functional sensitivity to token ordering. To distinguish structural organisation from distributional effects, we employ permutation-based null models that preserve marginal statistics while disrupting sequence structure. The results indicate that the learned representations exhibit behaviour consistent with non-trivial organisation beyond what can be explained by marginal statistics alone. This work demonstrates the feasibility of using discretisation to establish a testable framework for assessing structural organisation in learned representations, providing a principled pathway toward intrinsic structural interpretability beyond black-box modelling in PPG-to-ABP translation.

Publication Type:	Conference or Workshop Item (Paper)
Additional Information:	© 2026, the authors. This is the accepted manuscript of an article published by Springer. Please refer to the publisher’s terms and conditions for information on re-use.
Publisher Keywords:	Structural Interpretability, Discretisation, PPG-ABP mapping
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science R Medicine > RC Internal medicine
Departments:	School of Science & Technology School of Science & Technology > Department of Engineering
SWORD Depositor:	Symplectic Administrator

Export

Downloads