Georgia Institute of Technology
Machine Learning Across the Critical Care Continuum: Sensing, Prediction, Treatment, and Faithful Explanation
Abstract
dc:description.abstractThis dissertation develops machine learning methods across the critical care continuum, moving from physiological sensing to septic shock prediction, hemodynamic treatment recommendation, and faithful explanation. The central goal is to extract clinically meaningful information from physiological and clinical data while preserving interpretability and trustworthiness in high-stakes care. Aim 1 presents a subject-independent tidal volume estimation framework using single-lead electrocardiogram and tri-axial seismocardiogram signals from a wearable chest patch. By combining amplitude, frequency, and baseline-wander respiratory surrogates with random forest regression, the model achieved an RMSE of 181.45 mL and Pearson correlation of 0.61 without per-subject calibration, demonstrating the feasibility of calibration-free respiratory sensing from unobtrusive cardiac-mechanical signals. Aim 2 develops two septic shock prediction frameworks. A parsimonious waveform-based model using heart rate, heart rate variability, pulse arrival time, and mean arterial pressure achieved an AUROC of 0.83 at the 6-hour prediction horizon using only a five-minute look-back window and no laboratory data. A broader multimodal model integrating demographics, laboratory values, vital signs, and waveform features achieved AUROC values of 0.89 and 0.87 on internal and external validation, respectively. Across both approaches, blood-pressure-related physiology emerged as the dominant predictive signal. Aim 3 introduces XHemo, a framework for hemodynamic treatment recommendation and faithful explanation in septic shock. An offline reinforcement learning policy using CQL+AWAC with an axial transformer state encoder learned fluid and vasopressor strategies from 5,654 septic ICU trajectories. Because retrospective off-policy estimates do not establish prospective clinical superiority, the main contribution is the attribution-grounded explanation framework. By conditioning large language model explanations on Integrated Gradients feature attributions, XHemo reduced unsupported clinical claims by 57–79% and improved feature coverage by 41–70 percentage points. Together, these studies show how machine learning can support sensing, prediction, treatment, and explanation in critical care when models are developed as connected and evidence-grounded tools rather than isolated predictive systems.
Degree
thesis:*- Name thesis:degree_name
- Machine Learning, 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
-
- Soliman, Moamen
- Advisor dc:contributor.advisor
-
- Inan, Omer T.
- Committee members dc:contributor.committeemember
-
- Navuduru, Rishitha
- Ploetz, Thomas
- Anderson, David
- Clermont, Gilles
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82168