A Fairness- Aware and Computationally Efficient Deep Learning Framework for Blood Pressure Estimation

This thesis proposes a fairness-aware and computationally efficient deep learning framework for calibration-free cuffless blood pressure estimation using multimodal ECG and PPG signals. A unified benchmarking pipeline is first developed through signal-to-image transformation and a hybrid Joint 3D-CNN and Vision Transformer model, achieving strong agreement with AAMI and BHS clinical standards on the MIMIC-II dataset. The study then analyses the trade-off between predictive accuracy and deployment feasibility in wearable environments. Finally, a Dual-Stream Multi-Scale Temporal Network with contrastive fairness learning is introduced to reduce demographic performance disparities across age and gender. The proposed framework enables accurate, generalisable, and ethically responsible continuous cardiovascular monitoring for real-world digital health applications.