Files and links (1)
Thesis_Zengchenghao3.71 MB
Embargoed Access, Embargo ends: 06/29/2028
Dissertation
Information sharing for sequential data analysis based on Gaussian process and mixed effects models
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2026
Abstract
Sequential and longitudinal data arise in many scientific and engineering applications where observations are collected repeatedly over time, across related processes, or from heterogeneous subjects. A central statistical challenge in such settings is how to borrow information across related but non-identical units to improve prediction and inference when data are limited, while still preserving meaningful heterogeneity. This dissertation studies structured information sharing for sequential data analysis and develops three methodological contributions that implement this idea at two complementary levels: sharing information across related processes, and sharing information across heterogeneous subjects.
At the process level, the first contribution develops a real-time transfer active learning framework for functional regression and prediction based on a tailored multi-output Gaussian process. The framework addresses the cold-start problem by transferring information from data-rich source processes to a data-scarce target process. A convolution-process-based covariance structure is introduced to characterize nonlinear cross-process dependence and to support interpretable information sharing at the functional-relationship level, and an iterative Bayesian updating procedure aligns real-time posterior inference with the iterative nature of active learning. Theoretical analysis establishes monotone improvement of the integrated mean squared error objective under the proposed framework. The effectiveness of the framework is demonstrated on numerical studies and on two real case studies in electrochemical impedance spectroscopy and reduced graphene oxide field-effect transistor sensor calibration.
At the subject level, the second contribution develops a mixed-effect hidden Markov model for intensive longitudinal nurse fatigue data collected through a 14-day ecological momentary assessment study of 675 nurses, with 596 nurses contributing usable longitudinal fatigue records. The model represents fatigue progression through latent states and allows transition probabilities to depend on both observed covariates and subject-specific random effects, thereby supporting individualized transition assessment across a heterogeneous nursing population while borrowing strength across nurses through shared population-level structure. The analysis identifies poor sleep quality, night-shift work, age, and adequate time away from work as key predictors of fatigue transitions, and yields individualized transition matrices that can be used to inform scheduling and intervention strategies.
The third contribution addresses a methodological gap in the modeling of multivariate ordinal time series that is not resolved by existing approaches: how to jointly capture continuous latent-state dynamics, ordered categorical observations, and systematic differences in how subjects use the response scale. To address this gap, the dissertation develops the Random Effects Ordinal State-Space Model, which combines a multivariate continuous latent-state process with subject-specific dynamic thresholds initialized through random effects. This design recovers within-category variation at the latent level even when only coarse ordinal observations are available, and explicitly separates latent-state heterogeneity from heterogeneous use of the response scale across subjects. Estimation is carried out through a Monte Carlo expectation-maximization algorithm with sequential Monte Carlo smoothing, and theoretical results are established for identifiability, local convergence of the MCEM updates, and particle approximation consistency of the one-step-ahead predictive distribution. On the nurse fatigue application and case-calibrated simulation studies, the proposed model yields improved predictive performance relative to established baselines, with the largest gains in the data-scarce settings most relevant to clinical decision-making.
Taken together, these three contributions show that structured information sharing, whether implemented through a covariance structure across related processes or through random effects and random thresholds across heterogeneous subjects, can improve prediction and inference in sequential data settings while preserving the forms of heterogeneity most relevant to each application, and that the largest benefits arise precisely in the data-scarce regimes where traditional single-unit modeling is least reliable.
Details
- Title: Subtitle
- Information sharing for sequential data analysis based on Gaussian process and mixed effects models
- Creators
- Zengchenghao Xia
- Contributors
- Chao Wang (Advisor)Yong Chen (Advisor)Xin Zan (Committee Member)Amany Farag (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Industrial Engineering
- Date degree season
- Spring 2026
- Publisher
- University of Iowa
- Number of pages
- x, 128 pages
- Copyright
- Copyright 2026 Zengchenghao Xia
- Language
- English
- Date submitted
- 04/28/2026
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references (page 119-128).
- Public Abstract (ETD)
Many engineering and health applications collect data repeatedly over time, such as sensor measurements from battery testing, current–voltage measurements from reduced graphene oxide field-effect transistor sensors, and repeated fatigue ratings from nurses during work shifts. A common challenge is that the target process or person often has limited observations, while related processes or related people may contain useful information. This thesis studies structured information sharing, which means using related data to improve prediction and interpretation while still allowing each process or person to be different.
The first part develops a real-time transfer active learning framework based on a tailored multi-output Gaussian process. This method uses data-rich source processes to help learn a data-scarce target process, which is useful when experiments are costly or only a few measurements can be collected. The second part develops a mixed-effect hidden Markov model for nurse fatigue data. This model describes fatigue as transitions among low, moderate, and high fatigue states, and studies how sleep quality, work shift, age, and time away from work affect fatigue worsening or recovery. The third part develops a Random Effects Ordinal State-Space Model for repeated ordinal ratings. This model represents the underlying fatigue level as a continuous process while also accounting for differences in how people use rating scales.
Together, these methods show that sharing information in a structured way can improve prediction when data are limited. The results can support more efficient sensor testing and more individualized fatigue risk assessment.
- Academic Unit
- Industrial and Systems Engineering
- Record Identifier
- 9985177072602771
Metrics
1 Record Views