Personalized Blood Pressure Estimation With a Topological Prior Vector From Minimal Per-Subject Data

Abstract

Cuffless blood pressure (BP) estimation from photoplethysmography (PPG) signals using deep learning (DL) has attracted significant attention for continuous, noninvasive health monitoring. However, traditional machine learning relies on handcrafted features that capture only limited aspects of the signal. While DL can learn richer representations from data, it often requires large-scale training to achieve stable generalization across subjects and conditions and produces models with limited interpretability, hindering clinical adoption in safety-critical applications. To address these issues, we propose a framework that integrates subject-specific topological priors derived from persistent homology of PPG morphology. We construct a 6-D topological prior vector (TPV) that summarizes per-beat global morphology and inject it into a dual-branch CNN via TPV-conditioned temporal gating, emphasizing waveform segments that are informative for BP. We evaluate our model on the MIMIC-III PPG dataset. Even when trained from scratch using only a small number of subject-specific segments, it achieves mean absolute errors (MAEs) of ${3}.{82} {\,}~\pm {\,}~{0}.{53}$ mmHg for systolic BP (SBP) and ${2}.{63} {\,}~\pm {\,}~{0}.{41}$ mmHg for diastolic BP (DBP), demonstrating strong performance under limited-data conditions. Our approach embeds morphological priors tailored to each subject. This design significantly improves both data efficiency and personalization, enabling more practical BP estimation in real-world scenarios.