Stabilizing PPG-based Blood Pressure Estimation through Global-to-Personalized Learning

This study investigates the limitations of subject-dependent models for PPG-based blood pressure estimation and proposes a population-level pre-training and fine-tuning strategy. Due to limited subject-specific data, conventional models show high inter-subject variability and unstable performance. To address this, a population-level model was first trained on stable subjects and then fine- tuned on subjects with unstable estimations. The proposed approach reduced the mean absolute error of systolic blood pressure (SBP) from 15.79 to 4.85 mmHg and diastolic blood pressure (DBP) from 4.98 to 3.14 mmHg. These results demonstrate significantly improved estimation stability compared to purely subject-dependent training.