2026-05-19T02:51:52Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-2018602025-07-17T19:39:31Zoai_pmh:alloai_pmh:repo_items

201860 https://hdl.handle.net/2286/R.2.N.201860 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2025 133 pages Doctoral Dissertation Academic theses en Cai, Fulin Wu, Teresa Pedrielli, Giulia Lockhart, Thurmon Berisha, Visar Arizona State University Partial requirement for: Ph.D., Arizona State University, 2025 Field of study: Computer Engineering The increasing burden of chronic diseases and aging populations calls for scalable, efficient, and accessible diagnostic solutions. Biomedical sensors enable continuous monitoring of physiological signals, offering rich data for early disease detection and timely intervention. When combined with deep learning (DL), these signals can be transformed into automated, non-invasive diagnostic tools. However, two critical challenges persist in biomedical signal analysis: variability across frequency—particularly in multi-band signals with distinct power levels—and variation over time, reflecting dynamic physiological states and contextual changes. This dissertation proposes four DL-based frameworks that address these challenges across two categories of biomedical sensing tasks: (1) snapshot-based detection involving multi-band signals with varying power scales, as in gait abnormality detection and hemodynamic scenario classification; and (2) temporally dynamic detection, as required in sleep disorder diagnosis. To address frequency-related variability, a Band-Dependent Learning (BDL) framework is developed for Alzheimer’s disease and related dementia (ADRD) risk assessment using gait signatures from micro-Doppler radar. A digital twin simulation engine is introduced to generate realistic gait data exhibiting ADRD-specific abnormalities. Building on BDL, an enhanced BDL (E-BDL) framework is proposed to further improve classification and interpretability. E-BDL is also applied to radar-based vital sign monitoring for hemodynamic scenario classification by capturing cardio-respiratory variability across frequency bands. To address temporal variation, two novel frameworks are introduced for obstructive sleep apnea (OSA) diagnosis. The temporal proximity contrast learning framework models temporal deltas and relationships among consecutive electrocardiogram segments to mitigate performance degradation due to contextual variability in sleep. For fine-grained OSA event delineation, a per-second detection framework based on sleep sound recordings is proposed. This approach integrates temporal anchor contrastive learning and synergistic distillation to improve event localization and the estimation of key clinical metrics, including apnea-hypopnea index and mean apnea duration. Experimental evaluations across real-world and simulated datasets demonstrate state-of-the-art performance, robustness, and generalization of all proposed approaches. Overall, this dissertation presents DL-based methodologies that significantly advance biomedical signal analysis by addressing fundamental spectral and temporal challenges in clinical diagnostics. Computer Engineering Advanced Deep Learning Approaches for Enhanced Healthcare Diagnosis Using Biomedical Signal Data