DDCCR-Net: A Dual Dynamic Convolutional and Correlation-Based ResNet Model for Blood Pressure Prediction

This study aims to explore a cuffless approach for estimating systolic blood pressure (SBP) and diastolic blood pressure (DBP) using photoplethysmogram (PPG) and electrocardiogram (ECG) signals. To this end, we propose a novel framework based on dual-modal dynamic convolutions and correlation-based residual networks (ResNets)—DDCCR-Net. The framework first extracts low-level features from the PPG and ECG modalities using dynamic convolution and a channel-wise adaptive mechanism and combines dual-modal dynamic convolutions with learnable weights. This allows the model to effectively capture the complex dynamic variations in physiological signals. Then, a ResNet is employed to facilitate better feature propagation in the dynamic convolution module, enabling the model to capture higher-level dependencies in the signals and generate more expressive features. Finally, a correlation module is introduced to further learn interfeature relationships, enhancing the accuracy of the predictive features, which are then utilized for the final predictions. The proposed framework was validated on the VitalDB dataset, with experimental results demonstrating superior performance in both SBP and DBP estimation tasks (SBP MAE ± SD: $3.91~\pm ~5.96$ mmHg, DBP MAE ± SD: $2.28~\pm ~3.79$ mmHg), outperforming existing baseline models and meeting international blood pressure (BP) measurement standards such as those of the Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS). Furthermore, fine-tuning techniques were applied to improve model performance, and its cross-dataset generalization capability was assessed using the MIMIC-III dataset. Overall, the proposed DDCCR-Net framework provides a highly accurate and versatile solution for cuffless BP estimation.