He, Shan2025-11-112025-11-112025-11-11http://hdl.handle.net/10393/51027https://doi.org/10.20381/ruor-31500Cuffless Blood Pressure (BP) estimation using Electrocardiogram (ECG) and Photoplethysmogram (PPG) features is a promising alternative to cuff-based devices for continuous BP monitoring. However, its practical deployment faces several critical challenges: Respiration Modulation (RM) introduces confounding fluctuations into both physiological features and BPs, degrading estimation accuracy and trend tracking; existing models often fail to accurately track BP trends, which is crucial for long-term monitoring; poor model generalization to new subjects leads to performance degradation; personalization strategies are inefficient, typically requiring a large number of calibration measurements; and evaluation metrics are inadequate, primarily focusing on pointwise accuracy while neglecting trend tracking. To address these limitations, this thesis proposes a comprehensive framework. First, the often-overlooked impact of RM on features and BP is mitigated using Variational Mode Decomposition (VMD) for pre-processing, enhancing accuracy and trend tracking. Second, a novel lightweight estimation model, TCN-SENet-KAN, coupled with a customized shape oriented Soft Dynamic Time Warping (sDTW) loss function, is proposed to improve BP trend tracking and model generalization. Third, a robust three-step personalization framework is developed, leveraging Domain Adversarial Neural Network (DANN) pre-training for subject-invariant features, followed by sDTW-guided fine-tuning and Procrustes analysis for post-hoc alignment. Finally, a composite similarity metric is introduced to jointly assess both distance and trend accuracy for continuous monitoring scenarios. The proposed framework achieves Mean Absolute Error (MAE) of 4.34 mmHg for SBP and 2.70 mmHg for DBP using only two calibrations, meeting the IEEE 1708a-2019 standard for continuous BP monitoring and qualifying as “Grade A” performance. Furthermore, the method demonstrates robust BP trend tracking with median Pearson correlation coefficients of 0.781 (SBP) and 0.603 (DBP) across 32 subjects from the BCG BP dataset. Overall, this comprehensive framework significantly advances the practical viability of cuffless BP estimation for real-world applications.enAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Cuffless Blood Pressure EstimationBlood Pressure Trend TrackingBiomedical Signal ProcessingDeep LearningTime Series AnalysisThesis