Enhancing ICU Length of Stay Prediction Using Patient-Specific Factors; Legacy Challenges, Iterative Solutions: A Socio-Technical Analysis of EHR Integration and AI Potential

Innovation in healthcare is exciting—but never simple. Growing up with a parent working in healthcare, I witnessed how new technologies shape patient care and the daily realities of those on the front lines. This duality inspired my work on two projects that explore the intersection of technology and practice in healthcare. My technical project focuses on developing a neural network model to predict ICU length of stay (LOS). This model analyzes the relationship between patients, providers, and resources to help hospitals potentially rework their systems and optimize these relationships in order to minimize LOS. My STS research shifts the focus from technical optimization to understanding the broader sociotechnical implications of healthcare technology. It examines the adoption of Electronic Health Records (EHRs), exploring their influence on clinician workflows. Using Geels’ Multi-Level Perspective framework, I identified how implementation strategies shaped outcomes for providers and patients alike. The relationship between the projects lies in my model serving as a technical tool to optimize healthcare operations, while my STS research examines past implementations of similar tools, like EHRs, and how their success was influenced by human and systemic challenges. Without incorporating STS analysis, new healthcare solutions risk falling short of meaningful impact. The technical portion of my thesis focuses on building a neural network to predict ICU LOS, a critical issue in hospital resource management. Using the MIMIC-IV database—a comprehensive, de-identified dataset of hospital patient records—I trained the model to identify key factors like vitals, demographics, and clinical histories that influence LOS. The focus on the ICU addresses a gap, as existing models often generalize predictions across hospital settings. Additionally, the model integrates relationships among patients, providers, and resources—rarely used together in predictions. Improving ICU prediction accuracy by analyzing provider-patient-resource relationships together, rather than in isolation, is a step towards generalizing these results into concrete tools for healthcare planning. In my STS research, I analyzed EHRs through Geels' MLP framework to understand how implementation decisions shaped their outcomes. Through three EHR integration case studies—Vanderbilt University Medical Center, Mayo Clinic, and Cedars-Sinai—I explored how technological innovations, institutional practices, and external pressures interacted to influence the success of EHR adoption. Unlike previous studies that often focused on isolated technical issues or the challenges faced by individual institutions, my approach took a sociotechnical systems perspective, examining the broader context in which these technologies were deployed. My findings highlight the importance of ongoing stabilization at the niche level, where technology directly interacts with users. Successful integration relies on continuous adaptation to meet the unique needs of each institution. Stabilizing EHR use through iterative improvements and feedback loops ensures smoother adoption within the broader regime. This approach helps prevent new healthcare technologies, like AI, from overwhelming clinicians or disrupting workflows, enabling them to integrate more effectively without adding to clinician strain and ensuring they enhance, rather than hinder, healthcare operations. Reflecting on my work, I see how my technical and STS projects continuously shaped each other. When developing the neural network, I initially focused on prediction accuracy without considering how the relationships I was centering the model around—provider, patient, and resource—vary greatly across hospitals. This realization pushed me to examine how previous tools, like EHR systems, accounted for these differences—or failed to. Similarly, my STS research revealed that successful tools require iterative adaptation to meet the unique needs of different healthcare systems. This directly influenced my technical project, leading me to design a more interpretable model that could accommodate diverse hospital environments and improve trust in its predictions. With the challenging back-and-forth between my technical and STS work, I could not help but wonder how scientists have ever managed to create tools that truly thrive in dynamic systems that are constantly shifting and evolving. Thinking of healthcare in this way made me recall Vitruvius and his idea that early shelters, born out of the discovery of fire, were simple and adaptive at first, but over time evolved into more complex, life-like structures. Just like those early shelters were not truly considered architecture until they were improved to meet people’s needs, the technologies we create must keep evolving to fit the systems they are meant to serve. Centering the development of these tools around STS analysis keeps the ethical responsibility of incorporating outside perspectives at the forefront. While this approach may complicate or lengthen the process, it is essential for ensuring that the new technologies not only function well but also align with the needs and realities of the environments they are designed for. This synthesis would not have been possible without the guidance and support of Professor Kathryn Neeley. Her teaching shaped my understanding of sociotechnical systems and encouraged me to think critically about each component of my work. I also wish to thank the UVA Biocomplexity Institute and Professor Anil Vullikanti for providing the resources and infrastructure necessary for my technical project. He often had to pull me out of the weeds as I obsessively mapped every intricate relationship in my model, reminding me to think about how it would actually work in the messy, big-picture world of healthcare.