Mutual transfer learning for cuff-less blood pressure estimation using photoplethysmography-based visibility graphs

Cuff-less blood pressure (BP) estimation is critical to cardiovascular disease prevention and management. Photoplethysmography (PPG)-based monitoring technology offers advantages over cuff-based devices, including portability, lower power consumption, and faster measurements. However, current deep learning methods for BP estimation from PPG signals are limited by their analysis only from one-dimensional perspectives, failing to exploration of the underlying physiological patterns of cross-domain visual representations based on higher dimensional perspectives. Furthermore, conventional knowledge distillation techniques necessitate unidirectional knowledge transfer from pre-trained models, rendering it challenging to obtain feedback on the learning state of small networks for optimizing and adjusting the training process. This is inadequate for acquiring a profound understanding of the intricate mapping relationship between PPG and BP values. Therefore, this work presents a novel Transformer-based mutual transfer learning framework (MTL) that estimates BP values from phase-space reconstructed PPG signals using a multi-field complementary approach. The proposed MTL method leverages four phase-space reconstruction techniques to convert PPG signals into visibility graphs (VGs) that provide rich time-variant information. Furthermore, the joint optimization strategy with multiple losses of the soft label and structural knowledge learning enables us to transfer pre-trained knowledge from heterogeneous Transformer models and obtain cumulative multi-field complementary VG features during the fine-tuning process. We evaluate our MTL on three datasets of 1375 subjects using a subject-wise data-splitting paradigm based on five-fold cross-validation, achieving a state-of-the-art performance with estimation errors of 0.50 ± 4.94 millimeter of mercury (mmHg) and 0.21 ± 2.63 mmHg for systolic and diastolic BP, respectively. Our proposed end-to-end MTL offers a computationally efficient solution and elegant generalization ability for BP estimation using PPG-based VGs, providing a novel and innovative approach to BP monitoring that can advance cardiovascular disease prevention and management.