TPF-BP: A Three-Stage Physiology-Guided Self-Supervised Transfer Learning Framework for Personalized Blood Pressure Measurement.

Personalized and continuous cuffless blood pressure (BP) monitoring using photoplethysmography (PPG) is a critical requirement for next-generation wearable healthcare systems. Despite recent progress, existing PPG-based learning models exhibit limited generalization and personalization across individuals, hindering robust real-world deployment. To overcome these limitations, we propose TPF-BP, a novel self-supervised transfer learning framework that integrates physiological priors into representation learning. Specifically, a physiology-guided masked autoencoder is employed for self-supervised pre-training to enhance latent feature extraction from PPG signals and their derivatives, followed by a homologous triple-branch model for sequential supervised pre-training and personalized fine-tuning. Extensive experiments demonstrate that four base models trained with TPF-BP reduce the mean absolute errors of systolic and diastolic BP estimation by an average of 7.66% and 5.78%, respectively, in few-shot transfer scenarios across three public datasets and one self-collected dataset. Additionally, we publicly release CRHE, our self-collected multi-wavelength, dual-24h PPG-BP dataset, along with its benchmark results. The dataset features circadian rhythm annotations and was collected using wearable devices and ambulatory BP monitors. Together, TPF-BP and CRHE provide essential methodological and data resources to advance intelligent physiological modeling and personalized health monitoring. The dataset is publicly available at https://github.com/FDU-Bio-Monitoring/CRHE-Long-Term-PPG-BP-Dataset.