Abstract
Technical Project Abstract
Heart disease is the leading cause of death in the United States, and about 32 million of these cases are classified as heart failure with preserved ejection fraction (HFpEF). HFpEF is associated with increased mortality, reduced quality of life, and greater treatment difficulty. Cardiac magnetic resonance imaging (CMR) is a valuable tool for studying and diagnosing HFpEF, as it is a noninvasive imaging modality that enables quantitative assessment of myocardial structure, function, and tissue composition. Pre-clinical CMR in murine models is important for advancing research on HFpEF; however, it requires significant manual processing time, which can be prone to inconsistency across researchers. To solve this, my group and I created AI image analysis tools for two pre-clinical CMR modalities. Displacement Encoding with Stimulated Echoes (DENSE) is an image modality that assesses cardiac strain and diastolic function, and Cine quantifies cardiac structure.
First, we worked on data preprocessing to curate training data. We manually sorted through and rated CMR image quality and segmentation quality based on a rubric derived from the Likert scale. We used this data to develop our AI-based segmentation models. Both the DENSE and Cine models were based on 3-Dimensional U-Net architecture with a pretrained VGG16 backbone. We tested various hyperparameters and loss functions to optimize each model. The accuracy of these output segmentation masks was quantified using the Dice score, which is a standard metric for evaluation of image segmentation models on a scale of 0 to 1. Our goal was to have Dice scores of 0.85 or above for each model, and the DENSE and Cine models achieved Dice scores of 0.84 and 0.87, respectively. Finally, we integrated the models back into the existing data processing pipeline in the lab. This included creating a classification model to identify image slices for segmentation and file conversion to make sure the inputs and outputs were compatible with Medviso and MATLAB. Future work includes creating AI-based segmentation models for other CMR modalities used in the lab.
STS Project Abstract
The increasing integration of AI in clinical practice has introduced challenges regarding transparency, accountability, and trust. I examine the Epic Sepsis Model (ESM), a widely deployed predictive algorithm designed to detect early signs of sepsis in hospitalized patients, to analyze how sociotechnical factors influenced the performance and perception of this model. After deploying this model, external validation studies showed that the ESM missed about two-thirds of sepsis cases while also generating a high number of false positive alarms. Instead of attributing these problems only to technical limitations, I argue that the ESM’s failure is a result of its broader sociotechnical network.
I use Actor-Network Theory (ANT) as a conceptual framework to consider the interactions between human and nonhuman actors such as the model itself, ESM’s developers, clinicians, healthcare institutions, regulatory frameworks, and patients. I also draw on Cathy O’Neil’s framework of “weapons of math destruction” to explain how algorithmic opacity in black-box systems can produce harmful outcomes. Through my analysis, I show that the ESM’s influence in clinical practice was reinforced by gaps in AI regulation, as well as institutional pressures for efficiency and innovation. Clinicians remained legally and ethically responsible for patient outcomes despite their inability to evaluate the ESM’s predictions. Patients were the least powerful but the most affected group in the network. I explain the importance of addressing not only technical design but also the social, institutional, and regulatory aspects of AI systems. Improving clinical AI depends on increasing transparency, evaluation, and regulations throughout the sociotechnical network.
Connection Between Technical and STS Projects
My technical and STS projects both focus on AI in healthcare but from different perspectives. In the technical work, we developed AI-based segmentation models to improve efficiency and consistency of CMR imaging analysis in pre-clinical HFpEF research. We worked thoroughly to ensure accuracy, optimization, and integration into the existing workflow. My STS research is also about an AI model, the ESM, but this case study showed that technical design is not the only thing that determines success in clinical implementation. Sociotechnical factors such as transparency, accountability, and regulation shape how AI systems perform and how they are trusted in practice. Both of these projects together highlight that developing reliable medical AI tools requires thoughtful development and consideration of how the system interacts with clinicians, patients, and healthcare institutions. These tools have the potential to be incredibly useful, but we must learn how to responsibly develop and implement these tools first.
