Continuous noninvasive monitoring of blood pressure (BP) plays a crucial role in the early detection and management of hypertension and cardiovascular disorders. Photoplethysmography (PPG), widely used in wearable technologies, provides a practical and low-cost alternative to conventional cuff-based measurement systems. This study presents EstNet, a hybrid deep learning framework that integrates convolutional neural networks and gated recurrent units for the direct estimation of systolic and diastolic blood pressure from raw PPG signals. PPG data were obtained from the MIMIC-III waveform database, sampled at 120 Hz, and segmented into fixed-length windows. A structured preprocessing pipeline, including filtering, detrending, and normalization, was applied to enhance signal quality and reliability. Multiple machine learning and deep learning models were evaluated for comparison, and the proposed convolutional neural network-gated recurrent unit (CNN-GRU) model demonstrated the best overall performance. The model achieved a mean absolute error of 7.28 mmHg for systolic blood pressure and 5.01 mmHg for diastolic blood pressure using a subject-independent evaluation strategy. The findings indicate that hybrid convolutional-recurrent architectures effectively capture both morphological and temporal characteristics of PPG signals, enabling accurate cuffless blood pressure estimation. The proposed approach shows strong potential for integration into wearable systems for continuous and real-time health monitoring.

This study presents a comprehensive approach for non-invasive blood pressure estimation using photoplethysmography (PPG) signals by evaluating both classical machine learning and advanced deep learning models. The experimental results demonstrate that deep learning approaches consistently outperform traditional feature-based methods in terms of accuracy and reliability. Among all evaluated models, the CNN-GRU architecture achieves the best overall performance, recording the lowest error values and highest prediction consistency for both systolic and diastolic blood pressure. The model effectively combines convolutional layers for extracting morphological features and gated recurrent units for capturing temporal dependencies in PPG signals. It is also the only model to satisfy clinical standards for diastolic blood pressure estimation, highlighting its potential for practical applications. Overall, the findings confirm that hybrid deep learning architectures provide a strong foundation for cuffless blood pressure estimation and can significantly improve continuous health monitoring systems. Despite these promising results, several challenges remain before such systems can be widely deployed in real-world environments. The dataset used in this study is primarily based on intensive care unit recordings, which may not fully represent real-life wearable conditions. Factors such as motion artifacts, sensor placement variations, and environmental conditions can affect signal quality and model performance. Future work should focus on validating the model on diverse datasets collected from wearable devices to improve generalization. The integration of multiple physiological signals, such as electrocardiography and accelerometer data, may further enhance prediction accuracy and robustness. Additionally, advanced architectures, including transformer-based models, could improve the ability to capture long-term dependencies in PPG signals. Another important direction is the development of lightweight and energy-efficient models for real-time deployment on wearable and edge devices. Incorporating personalized and adaptive learning techniques may further improve accuracy by accounting for individual physiological differences.

[1]X. Zhang, Y. Li, L. Ling, J. Li, and C. Liu, “DDCCR-Net: A dual dynamic convolutional and correlation-based ResNet model for blood pressure prediction,” IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–12, 2025. [2]Y. Sivanjaneyulu, M. S. Manikandan, S. Boppu, and L. R. Cenkera-Maddi, “Quality-aware PPG-based blood pressure classification for energy-efficient trustworthy BP monitoring devices with reduced false alarms,” IEEE Access, vol. 13, pp. 132514–132536, 2025, doi: 10.1109/ACCESS.2025.3591977. [3]C. Botrugno, K. Dheman, P. Bonazzi, F. Dell’Olio, and M. Magno, “Optimized blood pressure prediction with a hybrid CNN–SVR model using multiwavelength PPG,” IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–12, 2025. [4]T. Xiang, W. Li, Z. Wang, and S. Liu, “Dynamic beat-to-beat measurements of blood pressure using multimodal physiological signals and a hybrid CNN–LSTM model,” IEEE Journal of Biomedical and Health Informatics, vol. 29, no. 8, pp. 5438–5451, 2025. [5]Y. Jia, Y. Zheng, Z. Wang, and Y. Li, “Non-invasive blood glucose based on dual-wavelength transmission PPG signal,” IEEE Sensors Journal, vol. 25, no. 7, pp. 11876–11884, 2025, doi: 10.1109/JSEN.2025.3539604. [6]H. Wang, Q. Xie, X. Li, Y. Qian, and G. Li, “Self-supervised contrastive learning for preprocessing PPG signals in blood pressure prediction,” IEEE Sensors Journal, vol. 24, no. 2, pp. 1637–1648, 2024, doi: 10.1109/JSEN.2023.3324700. [7]A. Schlesinger, S. Janosevic, G. Jakic, S. Antic, and M. Belic, “Multimodal cuffless blood pressure estimation using ECG, PPG, and continuous-time transformers,” IEEE Journal of Biomedical and Health Informatics, vol. 28, no. 3, pp. 1673–1685, 2024. [8]A. Hassan and M. Rafiq, “A deep temporal convolutional network for robust blood pressure estimation in real-world PPG signals,” Computers in Biology and Medicine, vol. 169, p. 107832, 2024, doi: 10.1016/j.compbiomed.2023.107832. [9]C. Ma, W. Deng, Q. Wang, W. Li, and G. Chen, “KD-Informer: A cuffless continuous blood pressure waveform estimation approach based on single photoplethysmography,” IEEE Journal of Biomedical and Health Informatics, vol. 27, no. 5, pp. 2219–2230, 2023, doi: 10.1109/JBHI.2022.3181327. [10]J. Kim, S. Bickel, and M. Jenkins, “Calibration-free BP estimation using multi-site PPG and a wave-based transformer,” IEEE Transactions on Biomedical Engineering, vol. 70, no. 11, pp. 3305–3316, 2023, doi: 10.1109/TBME.2023.3277809. [11]Y. Lee and J. Pan, “End-to-end hybrid CNN-GRU model for continuous blood pressure estimation from photoplethysmography,” Biomedical Signal Processing and Control, vol. 83, p. 104642, 2023, doi: 10.1016/j.bspc.2023.104642. [12]W. Cheng, Y. Sun, H. Liu, W. Li, and J. Chen, “Siamese CNN models for personalized continuous noninvasive blood pressure estimation,” IEEE Access, vol. 11, pp. 150233–150245, 2023, doi: 10.1109/ACCESS.2023.3339213. [13]F. El-Hajj, P. A. Kyriacou, and K. Budidha, “Motion-artifact-robust blood pressure estimation using adaptive filtering and deep recurrent networks,” IEEE Transactions on Biomedical Engineering, vol. 70, no. 4, pp. 1152–1162, 2023, doi: 10.1109/TBME.2022.3228125. [14]A. Ozkan, M. Karahan, and T. Yilmaz, “PPGNet: A lightweight deep learning model for real-time BP estimation in wearable devices,” Sensors, vol. 22, no. 19, p. 7421, 2022, doi: 10.3390/s22197421. [15]N. Liu, F. Xu, S. Wang, J. Lv, and Y. Li, “A multi-scaling residual network for cuff-less blood pressure estimation using PPG signals,” Biomedical Signal Processing and Control, vol. 77, p. 103748, 2022, doi: 10.1016/j.bspc.2022.103748. [16]Y. Liang, G. Li, Y. Zhang, Y. Zhao, and Q. Liu, “A new deep-learning model based on continuous wavelet transform for cuffless blood pressure estimation,” IEEE Sensors Journal, vol. 22, no. 3, pp. 2341–2351, 2022, doi: 10.1109/JSEN.2021.3131235. [17]Y. Li, W. Lin, K. Zhang, J. Lv, and J. Zheng, “End-to-end deep residual network for cuffless blood pressure estimation using single-channel PPG,” IEEE Sensors Journal, vol. 22, no. 15, pp. 14921–14931, 2022, doi: 10.1109/JSEN.2022.3177341. [18]M. H. Chowdhury, M. N. I. Shuzan, and M. B. I. Reaz, “A lightweight CNN model for blood pressure estimation using fingertip PPG signals,” Computers in Biology and Medicine, vol. 136, p. 104754, 2021, doi: 10.1016/j.compbiomed.2021.104754. [19]Q. Zhang, Z. Zhou, H. Wang, and Y. Liang, “A transformer-based network for cuffless blood pressure estimation using PPG signals,” IEEE Access, vol. 9, pp. 144574–144584, 2021, doi: 10.1109/ACCESS.2021.3122212. [20]S. S. Mousavi, M. Firouzmand, and D. Fattahi, “Blood pressure estimation using CNN–LSTM architecture from fingertip PPG signals,” Biomedical Signal Processing and Control, vol. 66, p. 102458, 2021, doi: 10.1016/j.bspc.2021.102458.