Date of Award

12-2026

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Biomedical Engineering and Sciences

First Advisor

Kunal Mitra

Second Advisor

Peshala Tibbotuwawa Gamage

Third Advisor

Christopher A. Bashur

Fourth Advisor

Linxia Gu

Abstract

Physiological vital-sign forecasting estimates future measurements based on recent temporal patterns and can support analysis of continuously recorded monitoring data. This study comparatively evaluated deep feedforward, recurrent, bidirectional recurrent, long short-term memory, and bidirectional long short-term memory architectures for one-step-ahead forecasting of peripheral oxygen saturation, heart rate, and pulse rate. Each model received consecutive observations of peripheral oxygen saturation, heart rate, pulse rate, respiratory rate, and age, while separate single-output models predicted the next value of the selected target.

Random and patient-wise data splitting were compared using identical input definitions, preprocessing procedures, model architectures, and training hyperparameters. The strongest architecture varied by target and splitting strategy, demonstrating that no single network consistently dominated all forecasting tasks. Patient-wise splitting removed patient- and temporal-window overlap, thereby providing a more rigorous assessment of model behavior on unseen patients, although its numerical effect remained architecture- and target-dependent.

The results are limited to a one-step forecasting horizon, at which consecutive physiological measurements are expected to be strongly autocorrelated. The dataset did not include a separate healthy reference cohort; therefore, healthy physiological values are discussed only as general context, and the findings are limited to continuous vital-sign forecasting within the analyzed patient cohort. Future work should evaluate persistence and statistical baselines, longer forecasting horizons, repeated patient-wise validation, healthy reference participants, and independent clinical datasets.

Comments

© Copyright 2026 Lavanya Vasavi Chittem Reddy

All Rights Reserved

 

 

 

 

 

__________________________________________________________________

The author grants permission to make single copies.

Download

Share

COinS