Blood Pressure Estimation from PPG for Wearable Devices: A Benchmark Study on Classical, Deep, and Transformer-Based Models

Accurate blood pressure (BP) monitoring is crucial for cardiovascular health monitoring and management. Traditional cuff-based methods, while reliable, are impractical for continuous tracking. This study explores a non-invasive approach to BP estimation using photoplethysmography (PPG) signals, leveraging classical machine learning (ML), deep learning (DL), and transformer-based models. A comprehensive benchmark analysis is performed across various datasets to assess the efficacy of these models. The research evaluates feature-based methods, like SVR, Decision Trees and Random Forests, alongside deep architectures, including CNN-based models such as CNN-LSTM, ResNet, UNet, ResUNet with attention, Transformer networks and Foundation Models. The models are tested on VitalDB dataset, obtaining for SBP and DBP respectively a lowest MAE of 5.33 and 3.42 mmHg in the case of calibration based test sets and 12.61 and 8.04 mmHg in the case of calibration free approach. Moreover the proposed methods are tested on smaller benchmark datasets, to verify how model performances change when applied to data of different sources. The results indicate that model accuracy depends on experimental conditions, including the measurement setup, the dataset used, and data acquisition methods, therefore specific use cases and data properties must be considered when choosing the BP estimation approach. The results indicate that DL models enhance BP prediction accuracy particularly on large datasets, when compared to traditional feature based ML methods on the same data. Furthermore, traditional ML methods achieve better results for smaller datasets, denoting the difficulties of adequately training DL on datasets of reduced size. Moreover, when incorporating ECG signals together with PPG, the BP prediction errors are notably reduced, highlighting the possibility of increasing accuracy with sensor fusion.