Abstract
Wearable and cuffless blood pressure (BP) monitoring hold great promise for preventive hypertension management, yet few studies have been validated under real-world and long-term conditions. In this study, we propose RACGBPNet, an efficient yet effective feature fusion framework for cuffless BP estimation and hypertension detection. The model leverages a two-stage fusion strategy: first, handcrafted and deep PPG representations are integrated via a self-attention mechanism to capture dynamic signal interactions; second, a simple gating module adaptively fuses signal features with demographic information. To mitigate severe class imbalance in hypertension detection, we further introduce a supervised contrastive learning scheme with a pretrained regression encoder. RACGBPNet was evaluated on two large-scale datasets, including a long-term smartwatch dataset spanning 30+ days in free-living conditions. On the public Aurora-BP dataset, RACGBPNet achieved state-of-the-art performance with mean absolute errors (MAEs) of 7.08 and 4.90 mmHg for systolic and diastolic BP, and Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.8627 and 0.8301 for systolic and diastolic hypertension detection under highly imbalanced data. Long-term validation confirmed that baseline calibration was sufficient to maintain compliance with AAMI standards for SBP up to 14 days and for DBP across the entire tracking period, striking a practical balance between accuracy and usability. However, the model exhibited significantly reduced variability compared with reference BP, indicating challenges in capturing abrupt BP changes. Overall, this work demonstrates the feasibility of long-term smartwatch-based BP monitoring and offers new insights into real-world deployment settings, particularly regarding calibration frequency and cross-dataset generalization.
Introduction
Hypertension is a major risk factor for cardiovascular disease, often termed a "silent killer" due to its asymptomatic nature. Traditional cuff-based monitoring, while accurate, is often disruptive and unsuitable for continuous daily tracking. Photoplethysmography (PPG)-based cuffless monitoring offers a non-invasive alternative, but transitioning these models from controlled clinical environments to real-world, long-term use remains a significant challenge. Key obstacles include the "Lab-to-Life" gap—where models fail over extended periods—and the inherent class imbalance in real-world data, where normal BP readings far outnumber hypertensive events.
The RACGBPNet Framework
To address these challenges, we developed RAGCBPNet (ResNet enhanced by Attention, Gating, and Contrastive learning). The architecture employs a sophisticated two-stage feature fusion strategy to maximize the information extracted from PPG signals and user demographics.
Two-Stage Feature Fusion
- Signal Fusion: A self-attention mechanism integrates 476 handcrafted features (e.g., pulse width, peak intervals) with deep features from a ResNet50 backbone. This allows the model to capture complex morphological interactions within the pulse waveform.
- Contextual Fusion: A gating module adaptively combines these signal features with demographic data (age, BMI, gender) and baseline BP. This mechanism allows the model to dynamically weight personal history against real-time sensor data.
Supervised Contrastive Learning
To improve hypertension detection in imbalanced datasets, we implemented a supervised contrastive learning scheme using TripletMarginLoss. This approach forces the model to learn embeddings that cluster hypertensive cases together while separating them from normal readings in the feature space.
Experimental Setup and Datasets
The framework was validated using two distinct datasets to test both accuracy and long-term stability:
- Microsoft Aurora-BP: A public dataset of 9,433 recordings from 626 subjects, used for benchmarking against existing state-of-the-art models.
- OPPO-BP: A new, large-scale dataset featuring 24,910 recordings from 136 subjects collected over 30+ days in free-living conditions.
Results and Long-term Stability
RAGCBPNet achieved state-of-the-art results on the Aurora-BP benchmark, with a Mean Absolute Error (MAE) of 7.08 mmHg for SBP and 4.90 mmHg for DBP. For hypertension detection, the model reached an AUROC of 0.8627 for SBP.
In the 30-day real-world study, the model demonstrated impressive stability. DBP estimation remained compliant with AAMI standards throughout the entire period. For SBP, accuracy was maintained for up to 14 days with a single baseline calibration. Analysis suggests that a 7-day calibration interval represents an optimal balance between user convenience and clinical accuracy for wearable devices.
Discussion and Limitations
While RAGCBPNet shows strong performance, it exhibits a "smoothing" effect, where it tracks general BP trends well but struggles to capture sudden, sharp fluctuations caused by acute stress or physical exertion. Furthermore, a "generalization gap" was observed when transferring models between datasets, with error increasing by approximately 1.5 mmHg. Future work will explore multi-modal sensor fusion (e.g., combining PPG with ECG) to better capture rapid vascular dynamics and improve cross-population robustness.
Summary updated