Title: Multi-modal Temporal Physiological Modeling Across Varying Time Horizons Using Dynamical Interactions

 

Date: Monday, July 6, 2026

Time: 11:00 AM - 1:00 PM ET

Location: TSRB523A + Remote 

Remote Meeting Link: https://gatech.zoom.us/j/99856255882?pwd=hgnvB6OKoETUl5k8DZGsXbioS7k43o.1

Remote Meeting ID: 998 5625 5882 | Passcode: 318520

 

Cem Okan Yaldiz

Robotics Ph.D. Candidate

School of Electrical and Computer Engineering

Georgia Institute of Technology

 

Committee

Dr. Omer Inan (Advisor) - School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. Thomas Ploetz - School of Interactive Computing, Georgia Institute of Technology

Dr. Vince Calhoun - School of Electrical and Computer Engineering, Georgia Institute of Technology

Dr. Reza Sameni - Wallace H. Coulter Department of Biomedical Engineering, Georgia Institute of Technology

Dr. Aaron Young – George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology

 

 

Abstract

This dissertation develops data-driven frameworks for cardiac-centered physiological analysis across multiple time horizons, leveraging multi-modal data acquired from wearable devices. At its core, the work emphasizes rigorous temporal modeling to capture the evolving dynamics of physiological states, enabling the design of predictive systems. The first contribution focuses on the early prediction of exertional heat stroke, a problem that inherently requires long-horizon inference to capture degradation in physiological state. By integrating deep learning with anomaly detection, the results show that multi-modal fusion achieves strong predictive performance, allowing the models to more faithfully represent the progression of physiological strain and stress. The dissertation then shifts to a short-horizon forecasting setting, targeting the precise timing of impending cardiac events such as aortic opening and closing. In this context, incorporating multi-modal inputs into a Kalman filter-based autoregressive framework is shown to significantly enhance accuracy, robustness, and computational efficiency. The work then examines the structural relationships among cardiac signals without target-based supervision to characterize a shared latent representation. By scaling the training data and primarily exploiting short windows of cardiac signals, the study demonstrates the effectiveness of structural self-supervised learning, particularly for distinguishing blood volume status levels. Finally, the scope is extended to cross-modal analysis of neural and cardiac signals, where multi-modal interactions are leveraged to enable the classification of major depressive disorder and preconscious responses. Taken together, this dissertation establishes that combining multi-modal fusion with principled temporal learning is key to modeling the complexity of physiological systems. Centered on cardiac data, the proposed approaches offer adaptable methodological frameworks, yielding both deeper physiological insight and robust, practical modeling strategies.