Holistic AI for Medicine, Explainability, and Optimization
Name
margaritis-geomar-phd-orc-2026-thesis.pdf
Size
6.89 MB
Format
Adobe PDF
Checksum (MD5)
3775e88ceba68addbf54193037e42780
Author(s)
Margaritis, Georgios
Advisor(s)
Bertsimas, Dimitris
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
The last decade has witnessed transformative advances in the field of Machine Learning (ML) and computation, fundamentally reshaping the landscape of scientific and industrial problem-solving.
This thesis leverages these breakthroughs to contribute to two parallel research agendas, centered respectively on Optimization and Healthcare.
In Optimization, we tackle the formidable task of solving general, non-convex optimization problems, where most traditional methods are unsuccessful.
The capacity to address such problems has been enabled, in part, by the substantial increases in speed and modeling power of Mixed Integer Optimization (MIO) solvers. By combining this power with advances in Machine Learning, we are able to design solvers that can address much more general Optimization Problems than existing approaches.
Concurrently and independently, in Healthcare, we focus on the real-world complexity of clinical data abstraction at scale.
Our objective is to develop and deploy models capable of achieving human-level accuracy for extracting structured, clinical information from diverse clinical sources.
Achieving this at a massive-scale, while maintaining high accuracy, generalizability to different hospitals and clinical explainability would be unattainable without the computational infrastructure and modeling power accrued over the past ten years.
However, recognizing that many critical application environments—such as hospital systems—often lack the computational infrastructure required to train heavy (language) models, we also develop methodologies that can maintain good performance with minimal computational cost.
This ensures that our proposed solutions are both effective and practically accessible in both high-resource and low-resource environments.
MIT Department
Massachusetts Institute of Technology. Operations Research Center
Sloan School of Management
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
