Abstract
Accurately predicting the length of stay (LOS) for patients in the intensive care unit (ICU) is vital for effective hospital resource management and patient care planning. My project investigates the use of machine learning (ML) models to forecast ICU LOS by leveraging the comprehensive, de-identified clinical data from the MIMIC-IV dataset. Unlike traditional approaches that focus solely on patient-specific factors such as demographics, vital signs, and lab results, this study incorporates engineered dependency features-capturing relationships between patients, healthcare providers, and resource utilization-to assess their impact on predictive accuracy. The methodology involves extracting and preprocessing time-series data, encoding categorical variables, and normalizing numerical features. Neural networks serve as the primary modeling approach, with hyperparameters optimized using automated frameworks. Model performance is evaluated using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and R-squared. Initial results show a test MSE of 14.883, MAE of 3.219, and a low R-squared of 0.043, indicating that the current model explains only a small portion of LOS variance. These findings highlight the complexity of ICU LOS prediction and the limitations of relying solely on patient-specific factors. I discuss limitations in feature representation, model complexity, and data quality, and propose future work involving advanced temporal models (such as RNNs or transformers), alternative ML methods (like XGBoost), and enhanced feature engineering. Ultimately, my research aims to improve predictive performance for ICU LOS, thereby supporting better resource allocation and patient outcomes in critical care environments.
Artificial intelligence (AI) is rapidly transforming higher education, offering opportunities for personalized learning, increased efficiency, and new modes of student engagement. My study examines university students’ perceptions of AI integration in their educational experiences and investigates the factors that shape acceptance or resistance to these technologies. Employing a mixed-methods approach—including surveys of 250 students, 15 semi-structured interviews, ethnographic observations, and historical analysis—the research provides a comprehensive view of student attitudes at the University of Virginia. The survey results reveal a generally positive outlook, with an average acceptance rating of 3.9 out of 5 and 68% of students believing that AI can improve their learning experiences. However, significant concerns persist: 64% of respondents distrust the transparency of data collection and use, and privacy concerns average 4.2 out of 5. Interviews and ethnographic observations highlight that while students appreciate the personalization and convenience of AI-driven tools, they remain skeptical about algorithmic bias, unreliable feedback, and the lack of human-like interaction. Actor-Network Theory (ANT) frames the analysis, emphasizing the dynamic interplay between students, AI technologies, institutional policies, and cultural norms. The study underscores the importance of transparent data practices, equitable access, and user-centered design to foster trust and maximize the benefits of AI in education. I offer recommendations for educators, policymakers, and developers to address ethical concerns and align AI integration with student needs. My findings contribute to ongoing discussions about the ethical, practical, and social implications of AI in higher education, advocating for responsible and inclusive technology adoption.
Both projects explore AI’s transformative potential in high-stakes environments (in this case, healthcare and education) while addressing ethical and practical challenges. My technical project focuses on improving ICU LOS prediction through ML, emphasizing dependency features and model optimization. In contrast, the STS project examines societal perceptions of AI, highlighting trust gaps and equity concerns. Despite differing domains, they share methodological rigor: the technical study uses neural networks and hyperparameter tuning, while the STS paper employs the ANT framework to dissect human-technology networks. A key intersection lies in their treatment of data transparency. The ICU model’s low R-squared value underscores the limitations of opaque patient data, paralleling students’ distrust in educational AI’s data practices. Both projects advocate for enhanced interpretability—whether through feature engineering in healthcare or policy interventions in education. Additionally, both highlight dependency interactions: the ICU study integrates provider/resource relationships, while the STS analysis frames AI tools as non-human actors shaping institutional workflows. Ethical considerations further unite them. The ICU research adheres to MIMIC-IV’s privacy protocols, mirroring the STS project’s emphasis on mitigating algorithmic bias and inequitable access. Ultimately, the technical project’s call for advanced temporal models aligns with the STS findings that AI’s efficacy hinges on user trust—a reminder that technological innovation must be grounded in sociotechnical accountability. Together, they illustrate how AI’s promise in specialized fields depends on balancing computational power with human-centric design.
