Date of Award
5-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Biomedical Engineering and Sciences
First Advisor
Peshala Priyadarshana Thibbotuwawa Gamage
Second Advisor
Christopher A. Bashur
Third Advisor
Mehmet Kaya
Fourth Advisor
Linxia Gu
Abstract
Continuous cuff-less blood pressure monitoring has the potential to transform how hypertension is detected and managed, but realising that potential requires sensor designs, validation protocols, and machine learning pipelines that are robust to real physiological variation. This thesis investigates the feasibility of estimating blood pressure from a single integrated sternal sensor combining electrocardiography and seismocardiography across a structured respiratory maneuver protocol within five physiologically diverse subjects. Three feature extraction strategies and three machine learning models were systematically compared across three complementary validation strategies, including a demanding activity-based holdout test that evaluated generalization from resting training conditions to physiologically opposite breathing perturbations. The most significant finding is the consistent superiority of windowed SCG features, which require no fiducial detection and no ECG timing information, over manually engineered beat-by-beat features in the majority of cases. This result raises the practical possibility that a single sternal accelerometer without ECG electrodes could serve as a sufficient sensing front end for a future wearable blood pressure monitoring system. The three validation strategies enabled a principled classification of model failures into two distinct types: recoverable distribution shift failures caused by activity-induced blood pressure range mismatches, and structural physiological ceiling violations caused by insufficient pressor response training exposure. A preliminary leave-one-subject-out analysis confirmed that personalized calibration is essential at the current cohort scale and that cross-subject generalization requires a substantially larger and more diverse cohort. Taken together the findings establish a clear feasibility foundation and a prioritized research road map toward the long-term goal of continuous ambulatory blood pressure monitoring in daily life.
Recommended Citation
Mohite, Shrushti Vinod, "Cuffless Blood Pressure Estimation from SCG and ECG Signals Across Respiratory Maneuvers: A Machine Learning Approach" (2026). Theses and Dissertations. 1644.
https://repository.fit.edu/etd/1644
Included in
Biomedical Devices and Instrumentation Commons, Other Biomedical Engineering and Bioengineering Commons
Comments
Copyrights held by author.