Abstract
My technical research project optimizes operating room (OR) scheduling within the UVA Health system by using machine learning and clustering to improve surgical duration predictions. My STS project analyzes how cultural values affect U.S. airlines' overbooking models and how the algorithms impact business and leisure travelers. While the projects aren’t directly connected by their topics, they are related by being optimization systems, efficiency-focused, and directly impacting their user groups, which include hospital staff and patients, as well as airline travelers. Both overbooking algorithms and our surgical duration model utilize predictive models to manage uncertainty, aiming to maximize efficiency and resource utilization.
The quality of OR scheduling depends on the accuracy of the surgical duration estimates. Current methods in the UVA Health system use a combination of historical moving averages, surgeon estimation, and the scheduling team’s expert prediction when necessary. However, those methods sometimes lead to inaccurate and variable results. My team and I developed machine learning models to improve the predictions. We relied on surgical data from July 2022 to June 2025 and used both feature engineering and clustering methods (TF-IDF and MiniLM) to improve the predictions when compared to the current scheduling method. We focused on two models: Extreme Gradient Booster (XGB) and Random Forest. Both models are ensemble learning models built on decision trees. The XGB model builds decision trees sequentially, each correcting the last, for high accuracy (Kavlakoglu & Russi, 2025). However, the random forest model builds many independent trees in parallel and averages their results for stability (Kavlakoglu, 2025). Both models significantly reduced the root-mean-squared error (RMSE), mean average error (MAE), and the number of surgeries running over. However, the XGB model performed best and reduced both MAE and RMSE by almost 60%. With these more accurate predictions, the scheduling team can make improved scheduling decisions and improve OR efficiency, reduce delays, and lower costs by reducing overtime. Our model is designed to work alongside the current scheduling team, not replace them, as we use the scheduled duration variable as a predictor for our model.
My STS research paper frames airline overbooking algorithms as sociotechnical systems that prioritize efficiency and profitability, leading to unequal impacts on passenger groups. This research uses a qualitative policy and document analysis of U.S. airline regulations, government reports, and academic literature to examine how overbooking systems operate in practice. Specifically, I analyze how airline overbooking models treat passengers as probabilistic units, using factors such as fare paid, loyalty status, and check-in timing to determine who is most likely to be denied boarding. These prioritization criteria are embedded within both airline practices and Department of Transportation regulations, which permit overbooking while structuring how airlines allocate risk among passengers. Business travelers are more likely to possess characteristics that these systems reward, such as frequent flyer status, higher fares, and lower price sensitivity, while leisure travelers are more likely to be price-sensitive, less loyal, and therefore more vulnerable to disruptions. This argument is supported through the analysis of airline revenue management literature, U.S. regulatory policies, and a case study of the 2017 United Airlines overbooking incident, which revealed how these systems become visible when they fail and generate public backlash. A brief comparison to South Korean airlines further illustrates this point, as carriers such as Korean Air and Asiana resolve overbooking issues before boarding begins, avoiding passenger removal altogether. This contrast highlights how cultural values, such as the United States’ emphasis on efficiency and individualism versus South Korea’s emphasis on collective harmony and uncertainty avoidance, influence how overbooking systems are designed and implemented. This argument was supplemented by Susan Leigh Star’s infrastructure theory, which explains that infrastructure is built on existing structures and often invisible until failure. Ultimately, the research demonstrates that overbooking algorithms are not purely technical tools, but systems shaped by U.S. cultural values of efficiency, profitability, and individualism, which normalize uneven distributions of risk and benefit across passenger groups.
These projects join together by investigating the relationship between efficiency and user impact, as overbooking policies decrease customer trust, and delays in surgery lead to decreased patient satisfaction and health outcomes. Jointly, the technical and STS projects illustrate how algorithms created to increase efficiency have social consequences, therefore highlighting the trade-offs between efficiency and consumer experience. My capstone project influenced my STS project by revealing how systems simplify complex human variability into data. Similar logic applies to the airline industry, where passengers are reduced to data points based on their fare, loyalty status, timing, etc. My capstone project helped me understand how algorithms operationalize priorities. My STS project influenced my capstone project by forcing me to think about who benefits from the optimization and what factors are excluded. It highlighted that models are not just technical tools, but they also impact users’ daily lives and experiences. Finally, it encouraged reflection on the ethical and practical implications of our models' use in the UVA Health system. By pursuing both projects together, I learned that efficiency-driven systems are powerful but not neutral. While optimization improves overall system performance, it also shapes how benefits and risks are distributed among users. This experience emphasized the importance of designing systems that consider not only technical performance, but also their broader human impact.
References:
Kavlakoglu, E. (2025, November 20). What is Random Forest?. IBM. https://www.ibm.com/think/topics/random-forest
Kavlakoglu, E., & Russi, E. (2025, November 17). What is XGBoost?. IBM. https://www.ibm.com/think/topics/xgboost
