Deep Learning To Classify Heart Failure with Preserved Ejection Fraction From Cardiac MRI Images; The Ethical Implications of Applying Large Language Models to Medical Practices

Author:
Le, Alexander, School of Engineering and Applied Science, University of Virginia
Advisors:
Earle, Joshua, EN-Engineering and Society, University of Virginia
Zhang, Miaomiao, EN-Elec & Comp Engr Dept, University of Virginia
Abstract:

Technical Project Abstract

This research constructs a deep learning framework for the diagnosis of Heart Failure with Preserved Ejection Fraction (HFpEF) and other cardiovascular conditions using multimodal cardiac magnetic resonance (CMR) imaging. we utilized extensive automated pipelines for image sorting, pre-processing, and a strong deep learning architecture to accurately distinguish among different cardiac pathologies, including HFpEF, Cardiac Amyloidosis, and Cardiac Sarcoidosis. Using the Heart Observations and Outcomes (HOO) database, our research aims to integrate cine imaging of several chambers, T1 imaging, Late Gadolinium Enhancement, and T2 mapping to create an influential system of classification.

The semi-automatic preprocessing pipeline combines manual region of interest identification and cardiac orientation with automatic techniques like ROI cropping, image rotation, histogram matching, and intensity normalization. In our baseline approach, we focused on short-axis cine images with the center slice over time frames in order to discriminate features. We simplified the classification task by converting multi-class labels into a binary one to differentiate between normal cardiac function and disease presence.

This work is part of the broader Heart Observations and Outcomes using Digitally Archived Twins Analysis (HOO DATA) project, whose goal is to create digital cardiac twins for precision medicine. Our computational technique is of great medical need because HFpEF affects approximately 3 million Americans but has no FDA-approved treatments due to its heterogeneity. By leveraging deep learning to identify significant imaging biomarkers that define unique HFpEF groups, this research serves to further diagnose and, ideally, tailor therapy for this prevalent and under-treated cardiac disease.

STS Project Abstract

My research analyzes the ethical impacts of the use of Large Language Models (LLMs) on the practice of medicine. Drawing on my literature reviews, qualitative interviews of doctors, and evaluation using the care ethics framework, I argue that LLMs are reshaping traditional care relationships in medicine. This shift is occurring in several ways: creating new layers between doctors and patients, changing access to and application of medical information, changing time devoted to patient care, and redistributing responsibility for care among healthcare workers. The impact is felt beyond direct patient care to clinical support systems, where LLMs are augmenting decision-making. I identified four key ethical concerns regarding LLMs in medicine: bias, hallucination, overreliance, and privacy issues. From a care ethics perspective, these concerns are reimagined as greater than technical errors, as they represent inherent failures in the ethical duty of caring for vulnerable populations. My dialogue with Dr. Hoke and Dr. Kieu highlights the current limited integration of LLMs into clinical practice, representing both technological caution and the priority of human care in relationships. Though LLMs are promising solutions to administrative burden and clinical decision support, more work needs to be done to reduce the potential harm that such systems may perpetrate, resulting in decreased quality of care relations. To achieve success in leveraging LLMs in healthcare, technological advancements should be complemented by the imperative norms of care ethics - in asserting that AI improves, and does not deteriorate, the human potential for responsive, attentive, and competent care.

Connection between Projects

My STS and technical projects share similarities in the field of artificial intelligence. In my technical project, I utilized deep learning models, a form of artificial intelligence, to create a model that is capable of classifying different heart conditions based on their cardiac magnetic resonance imaging scans. This closely relates to the issues that I discuss in my STS project: bias, hallucination, and overreliance. In my technical project, being aware of the bias that can be present in my dataset is important because ensuring that my training dataset captures a fair sample of the population ensures proper performance across a diverse population. Regarding overreliance and hallucination, I need to ensure that my model has proper oversight from a physician. By doing this, we can validate the model inputs to prevent any integration of hallucinations into the clinical decision-making process. My technical project also exemplifies what I analyze in my STS research: the growing intermediary of artificial intelligence in diagnosis that can potentially alter doctor-patient relationships. This support tool into clinical decision-making raises questions about responsibility, and who is accountable should my model misclassify a condition. Moreover, the care ethics framework from my STS project can evaluate whether my technical project can enhance provider’s capacity for attentive, responsive care. Both of my projects aim to optimize AI integration that supports, rather than replaces human judgement in healthcare.

Degree:
BS (Bachelor of Science)
Keywords:
Deep Learning, Cardiology, Large Language Models, Artificial Intelligence
Notes:

School of Engineering and Applied Science

Bachelor of Science in Computer Science

Technical Advisor: Miaomiao Zhang

STS Advisor: Joshua Earle

Technical Team Members:

Language:
English
Rights:
All rights reserved (no additional license for public reuse)
Issued Date:
2025/05/07