F-UTransBPNet A Transformer-Enhanced U-Net for Cuffless Blood Pressure Estimation Across Activity Scenarios

Abstract

Continuous blood pressure (BP) monitoring is essential for early detection and management of hypertension, a major risk factor for cardiovascular disease and death. Traditional cuff-based devices are not suitable for continuous or ambulatory monitoring because they can cause discomfort and only provide intermittent measurements. This has led to the development of cuffless BP estimation methods that utilize physiological signals like electrocardiography (ECG) and photoplethysmography (PPG). Existing deep learning techniques, such as convolutional neural networks (CNNs) and hybrid CNN–LSTM models, have shown promising results in controlled settings; however, they often face challenges in maintaining performance under changing physiological conditions. To address this issue, we introduce F-UTransBPNet, a hybrid U-Net–Transformer architecture that combines local feature extraction with long-range temporal modeling, further optimized through selective fine-tuning. The model was trained and tested on three complementary datasets: MIMIC (ICU patients), Dataset_Drink (water intake), and Dataset_Exercise (bicycle ergometry), covering both static and activity-based scenarios. Results show that F-UTransBPNet achieves mean absolute differences of 4.4 mmHg (SBP) and 2.2 mmHg (DBP) when compared to invasive references, meeting AAMI standards, and maintains strong correlations in activity datasets (PCC up to 0.82 for SBP). Minimal fine-tuning with just 10–20% of scenario-specific data restores performance across different domains, demonstrating the feasibility of calibration-light operation. These findings indicate that F-UTransBPNet offers an effective balance of accuracy, adaptability, and computational efficiency, supporting its potential as a reliable tool for cuffless BP monitoring in both inpatient and wearable healthcare settings.