Georgia Institute of Technology
Optimization for Interpretable Predictive and Prescriptive Modeling in Healthcare
Abstract
dc:description.abstractAs machine learning and artificial intelligence increasingly inform decision-making in high-stakes domains such as healthcare, there is a growing need for models that are not only accurate but also interpretable and trustworthy. This dissertation develops optimization-based methods for interpretable modeling in healthcare, addressing problems in disease prediction within maternal and perinatal health and medical decision-making in the context of chronic disease management. Motivated by the challenge of accurate risk estimation in maternal and perinatal health, Chapter 2 builds sparse and interpretable logistic regression models to predict a composite adverse perinatal outcome at multiple time points during pregnancy. Our findings enable clinicians to understand how risk factors for these adverse outcomes evolve over time, toward facilitating timely intervention. Chapter 3 builds on the work in Chapter 2 by addressing the complexities and disadvantages of modeling a composite outcome. Chapter 3 proposes the Composite Clustered Multi-Task Learning (CCMTL) problem, an optimization framework that simultaneously clusters individual outcomes into composite outcomes and learns predictive models for each cluster. We prove that CCMTL is NP-hard and derive both an exact non-linear mixed integer program and a scalable greedy heuristic algorithm to solve the problem. Through numerical experiments and a case study on maternal and perinatal health, we demonstrate that CCMTL can generate composite outcomes which exhibit improved prediction performance and reveal meaningful clusters of outcomes sharing underlying similarity. Chapter 4 explores interpretability in sequential decision-making by studying the trade-off between one-size-fits-all treatment guidelines and personalized care. Specifically, we formulate the Treatment Guideline Design Problem (TGDP), which simultaneously partitions a patient population into a pre-determined number of groups and designs an optimal treatment policy for each group. We characterize the structural properties of TGDP, proving that it is NP-hard and showing that population-level health outcomes are non-decreasing in the number of groups. To solve the problem, we develop exact mixed-integer linear programming formulations alongside heuristic solution methods. Our numerical experiments and case study on hypertension treatment planning demonstrate that TGDP can generate parsimonious treatment guidelines that achieve health outcomes within 1\% of optimal personalized policies while enhancing ease of implementation. Finally, Chapter 5 concludes with an overview of future research directions.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Industrial and Systems Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Lee, Sun Ju
- Advisor dc:contributor.advisor
-
- Garcia, Gian-Gabriel
- Committee members dc:contributor.committeemember
-
- Serban, Nicoleta
- Boulet, Sheree
- Xie, Weijun
- Steimle, Lauren
Subjects
dc:subject × 7Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82058