Compact PPG-Based Blood Pressure Estimation Using Hardware-Efficient Knowledge Distillation

Abstract

Continuous blood pressure (BP) monitoring from photoplethysmography (PPG) is compelling for wearable health applications, yet deep learning models for cuffless BP estimation are rarely evaluated under edge-deployment constraints. We propose a hardware-oriented knowledge distillation framework for compact PPG-only BP estimation, where deployable students are trained using ground-truth BP labels, teacher predictions, and feature-level supervision from either PPG-only or hybrid ECG–PPG teachers. Evaluated on PhysioNet PTT and VitalDB, knowledge distillation improves or maintains the accuracy–efficiency tradeoff of compact students, especially in datasets with minimal data. A compact 1-depth 16-channel Residual UNet student model maintained a deployable size of 20.7 k parameters with minimal penalty on average error. Hardware evaluation shows that the smallest, yet acceptable, student model fits within the VCU709 FPGA budget, while nRF5340 microcontroller unit deployment remains RAM-limited under full-window inference.