Abstract
Artificial intelligence (AI) technologies have become a clinical reality, with over 1,300 such tools FDA-approved as of 2026. In the healthcare space, the end goal of these tools lies not in technical performance but at the patient’s bedside. AI models are already able to segment cardiac images and identify anomolaies in X-rays faster than an expert physician, but technical prowess can mask the underlying ethical risks that come with implementing these powerful tools. My work navigates the intersection between these ideals, combining the development of technically advanced AI tools with an evaluation of the ethical impact these tools will have on the healthcare field. My technical project sits at the very beginning of the development cycle for AI technologies. My team and I created two different automated data analysis tools that leverage deep learning models to greatly reduce the time required to analyze cardiac magnetic resonance images (MRI) in preclinical models. For my STS research, I chose to study the opposite end of the development lifecycle, focusing on how AI tools are deployed safely and effectively in clinical settings. To accomplish this, I organized my research under the four pillars of medical ethics framework, specifically utilizing the pillars of beneficence, nonmalefience, and justice. This led to a better understanding of not only the pros and cons AI tools, but also how these impacts may differ between high and low resource communities.
The technical portion of my thesis produced two independent deep learning pipelines for preclinical cardiac MRI image analysis, automating the segmentation of both DENSE and Cine images. Previously, the manual segmentation of DENSE and Cine images took over an hour. By utilizing a 3-D U-Net model trained on manually segmented images generated by expert researchers, these pipelines are capable of generating a complete segmentation in under a minute. For the DENSE pipeline, this model was integrated into existing software developed by the lab, allowing researchers to seamlessly generate segmentations with the single click of a button. For Cine segmentation, we built a standalone analysis pipeline from scratch, applying both a classification model to identify the beginning and end of the heart beat and a segmentation model to recognize the heart at those timepoints (Figure 1).
In my STS research topic, I applied the four pillars of medical ethics framework to evaluate a variety of ethical concerns attached to the rapid expansion of artificial intelligence into the clinic, focusing specifically on diagnostic tools. My work confirmed that AI diagnostic tools are capable of matching or even surpassing the technical accuracy of physicians, while completing the analysis much more quickly. At the same time, reliance on these tools causes physician deskilling and susceptibility to incorrert suggestions, potentially putting patients at risk. My research revealed that when implemented correctly, AI tools are capable of closing the gap between high and low resource healthcare settings by alleviating issues caused by physician availability. However, the cost of AI diagnostic tools can be prohibitive for many of these same communities that would benefit the most. An important takeaway from my work is the need for more research into the clinical outcomes achieved by AI diagnostic tools, as opposed to the current focus on technical benchmarks. Additinally, more work studying the possibility of clinician deskilling needs to be done. To address the concerns with implementing AI tools in low resource settings, I argue that AI researchers should emphasize developing portable systems capable of plugging into both modern and lower resource settings, ensuring they are present in the spaces where they have the potential to make the largest impact.
By concurrently working on both the technical development of AI systems and an ethical assessment of these tools provided me with a better understanding of both problems. My STS research revealed an overreliance on technical metrics in assessing AI tools in healthcare. Because of my technical background, I am aware of necessity of these metrics during development, which can make it easy to view them as the endpoint when development is completed. However, there is a need to continue assessing AI models past this point to better understand how these tools impact patients.
