Blood Pressure Estimation Using Self-Attention Mechanism Built-In ResUNet on PulseDB: Demographic Fairness and Generalization

This study focuses on the estimation of systolic and diastolic blood pressure (SBP and DBP) from photoplethysmogram (PPG) and electrocardiogram (ECG) signals using deep neural networks (DNNs). We contributed to two significant areas. First, by implementing the self-attention mechanism to the ResUNet architecture improving the model’s ability to capture contextual dependencies. Second, using the PulseDB dataset—a large, preprocessed dataset based on MIMIC-III and VitalDB developed specifically for developing and benchmarking cuffless blood pressure (BP) estimations. This enabled us to both evaluate our model’s performance using a calibration-free approach and investigate the impact of gender and age on its accuracy. In calibration-based testing, our model achieved a mean absolute error (MAE) of 1.13 mmHg for DBP, outperforming current state-of-the-art models, and 4.45 mmHg for SBP, showing comparable accuracy. In calibration-free testing, the model recorded an MAE of 5.12 mmHg for DBP and 8.11 mmHg for SBP. Both the results meet the standards set by the Association for the Advancement of Medical Instrumentation, IEEE-1708, and the British Hypertension Society. Notably, our analysis also quantified, for the first time, the extent of overestimation associated with the calibration-based approach. In addition, we explored cross-dataset evaluation to assess generalization, applying fine-tuning to enhance performance. Importantly, stratified analysis highlighted the significant impact of age and gender on BP estimation, emphasizing the necessity for fairness in future models. This research advances the field by introducing a robust architecture and addressing dataset limitations, paving the way for more accurate and fair BP estimation methods.