Georgia Institute of Technology
Advancing Responsible AI in Clinical Decision Support Systems
Abstract
dc:description.abstractArtificial intelligence has become a powerful engine for clinical prediction, yet strong predictive performance alone does not make a model trustworthy, inspectable, or usable at the bedside. This dissertation advances Responsible AI for clinical decision support systems (CDSSs) by operationalizing three developer-controllable pillars—reliability, faithful interpretability, and translation—that together address the current grand challenge of turning high-performing AI into systems clinicians can trust, inspect, and actually use. Aim 1 establishes the reliability pillar by developing ensemble-based confidence calibration and uncertainty estimation methods for deep-neural-network-based respiratory disease screening. Through a conservative majority-vote calibration strategy (ENCL-DNN), the work demonstrates that predicted probabilities can be aligned with true correctness rates and that predictive entropy can separate confident from uncertain cases, enabling clinicians to trust, act on, or override model outputs. Aim 2 advances the faithful interpretability pillar by progressing from soft, inspectable intermediates toward hard, auditable evidence. Through concept bottleneck architectures (KindSleep) and discrete evidence search (Tree-of-Evidence), the work establishes that forcing predictions through clinically meaningful intermediate structures yields explanations that faithfully reflect actual model reasoning, not merely plausible post-hoc stories, while preserving or improving predictive accuracy. Aim 3 operationalizes the translation pillar by developing AirQuant, an automated pipeline that quantifies airway remodeling after orthognathic surgery across diverse craniofacial populations, including patients with orofacial clefts who are routinely excluded from published evidence. The results reveal that baseline anatomy, not procedure labels alone, drives remodeling, and that shape-based metrics capture changes that volume alone misses, demonstrating that translation requires patient-specific, equitable clinical decision support. Collectively, this dissertation demonstrates that Responsible AI in clinical decision support requires designing systems that produce calibrated confidence under uncertainty, route predictions through faithful and auditable evidence, and translate into equitable tools for real patients across real clinical workflows.
Degree
thesis:*- Name thesis:degree_name
- Electrical and Computer Engineering, PhD
- Department thesis:degree_department
- Electrical and Computer Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Nnamdi, Micky C.
- Advisor dc:contributor.advisor
-
- Wang, May Dongmei
- Committee members dc:contributor.committeemember
-
- Luo, Yunan
- Purnell, Chad
- Anderson, David
- Heck, Larry
- ORCID person:identifier.orcid
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81789