Temporal-Aware and Memory Fusion-Based Non-Invasive Continuous Blood Pressure Prediction

Aiming at the intermittent measurement limitations of traditional blood pressure monitoring methods and the insufficient feature fusion in existing deep learning models, this study proposes a hybrid model integrating Temporal Convolutional Network and Bidirectional Gated Recurrent Unit. The model captures multi-scale spatial features from photoplethysmography waveforms through TCN's dilated convolutions, models long-term hemodynamic temporal dependencies using BiGRU's bidirectional gating mechanism, and incorporates an attention mechanism for dynamic spatiotemporal feature fusion. Clinical dataset validation demonstrates that the model achieves MAE metrics of 2.59 mmHg and 3.32 mmHg for systolic and diastolic blood pressure prediction respectively, showing significant advantages over traditional machine learning methods and single deep learning models. This framework enables continuous blood pressure estimation with medical-grade accuracy while maintaining computational efficiency through synergistic learning of local waveform characteristics and global temporal patterns.