Georgia Institute of Technology
Hipiess: A Framework for Artificial Intelligence and Machine Learning in Clinical Informatics
Abstract
dc:description.abstractThis thesis addresses a fundamental challenge in modern healthcare artificial intelligence (AI): the disconnect between advances in machine learning (ML) and their effective use in real-world clinical settings. While recent progress in data integration, predictive modeling, and large-scale representation learning has enabled highly accurate models, these systems remain difficult to interpret, integrate, and deploy in ways that meaningfully support clinical decision-making. As a result, there exists a persistent gap between algorithmic capability and clinical impact. The core argument of this work is that this gap does not arise from deficiencies in any single component—data, models, or deployment—but from the lack of a unified framework that integrates these components into a coherent, clinically usable system. Existing approaches to data harmonization, such as extract, transform, and load (ETL) pipelines and standardized models, enable aggregation of heterogeneous data but remain static, inflexible, and poorly aligned with downstream modeling needs. Similarly, advances in machine learning, including transformer-based architectures and large language models (LLMs), provide powerful predictive capabilities but introduce significant challenges in interpretability. Although explainable artificial intelligence (XAI) methods such as feature attribution and post-hoc explanation techniques offer partial insight into model behavior, they often fail to align with clinical reasoning, are not actionable, and are rarely integrated into clinical workflows. Finally, deployment systems such as clinical decision support systems (CDSS) face challenges in scalability, usability, and integration, further limiting the translation of models into practice. To address these limitations, this thesis proposes the HIPIESS framework as a unified architecture for clinical machine learning systems. HIPIESS treats data processing, modeling, interpretability, and deployment not as isolated stages, but as interconnected layers within a single system. At the data layer, heterogeneous clinical inputs are continuously harmonized into a unified representation space. At the model layer, predictive functions are designed with interpretability as a first-class constraint, enabling explanations that are aligned with clinically meaningful concepts. At the system layer, outputs are integrated into scalable, workflow-aware decision support systems that provide actionable guidance. A key contribution of this work is the formalization of HIPIESS as a compositional framework, where clinical data (D) are transformed through harmonization H(D), modeled via predictive functions M(H(D)), interpreted through explanation functions E(M), and deployed through support systems S(M,E). This formulation highlights the interdependence of each component and emphasizes the importance of feedback between layers, enabling continuous adaptation and improvement. By shifting the focus from model-centric interpretability to system-level integration, this thesis advances a new perspective on clinical artificial intelligence: that interpretability, usability, and scalability must be co-designed rather than optimized independently. The HIPIESS framework provides a foundation for developing clinical machine learning systems that are not only accurate, but also interpretable, deployable, and capable of delivering real-world impact.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, PhD
- Department thesis:degree_department
- Computational Science and Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hornback, Andrew
- Advisor dc:contributor.advisor
-
- Wang, May Dongmei
- Committee members dc:contributor.committeemember
-
- Luo, Yunan
- Joyner, David
- Varma, Sashank
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82193