Background. Hypertension is a leading modifiable risk factor for cardiovascular disease, making continuous, comfortable blood pressure (BP) monitoring clinically important. Cuffless BP estimation using deep learning (DL) from Photoplethysmography (PPG) and Electrocardiogram (ECG) signals is a promising alternative to cuff-based devices, yet existing DL approaches often extract PPG/ECG morphology poorly, fuse multi-modal and multi-depth information inadequately, and are evaluated under protocols that are not always fully subject-independent. Objective. This thesis designs and evaluates MDMF-Net, a physiology-aware DL framework that explicitly models PPG/ECG morphology and fuses multi-modal, multi-depth information for accurate cuffless estimation of Systolic BP (SBP) and Diastolic BP (DBP). Methods. MDMF-Net comprises three components: (1) a morphological feature extraction network, pairing a Standard-SE-ResNet branch for PPG with a Phy-SE-ResNet branch for ECG that adds a derivative path and dual-path attention for R-wave morphology; (2) a Cross-Depth Fusion Pyramid Network (CDFPN), performing triple-gated cross-modal fusion, temporal alignment, and feature enhancement before pyramid-style cross-depth fusion; and (3) a multi-scale prediction network with a learnable, softmax-normalized weighting mechanism. The model was trained and tested on the UCI cuffless BP dataset (132,887 5-second PPG/ECG segments from 942 subjects) under a leakage-free, record-wise partitioning strategy. Results. MDMF-Net achieved Pearson correlations of 0.923 (SBP) and 0.80 (DBP); Mean Error ± Standard Deviation of 0.325 ± 7.062 mmHg (SBP) and −0.639 ± 5.336 mmHg (DBP); and Mean Absolute Errors of 5.075 mmHg (SBP) and 3.761 mmHg (DBP). Both parameters satisfy the AAMI standard and achieve BHS Grade A, a combination unmatched by any of six compared baselines. MDMF-Net led on nearly all SBP metrics, remained competitive on DBP, and used roughly six times less GPU memory and half the inference time of its closest DBP competitor. Ablation studies confirmed each of the three components contributes independently to performance. Conclusion. Physiology-aware morphological feature extraction, combined with principled cross-modal/cross-depth fusion and dynamic multi-scale prediction, yields a cuffless BP estimation model that is more accurate, more clinically compliant, and more computationally efficient than
