Margrét Bjarnadóttir Directory Page

There is growing interest in the application of machine learning models and advanced analytics to various healthcare processes and operations, including the generation of new clinical discoveries, development of high-quality predictions, and optimization of administrative processes. Machine learning models for prediction and classification rely on extensive and robust datasets, particularly for deep learning models common in health, creating an urgent need for large health datasets. Yet datasets can be insufficiently large due to the rapid evolution of diseases, such as coronavirus disease 2019 (COVID-19), rarity of disease, or the myriad obstacles to sharing and acquiring existing health data, including ethical, legal, political, economic, cultural, and technical barriers. Synthetic data provide a unique opportunity for health dataset expansion or creation by addressing privacy concerns and other barriers. In this paper, we review prior literature and discuss the landscape of machine learning models used for synthetic health data generation (SHDG), outlining challenges and limitations. We build on existing research on the state of the art in SHDG and prior broad explorations of the potential risks and opportunities for large language models (LLMs) in healthcare. We contribute to the literature with a focused assessment of LLMs for SHDG, including a review of early research in the area and recommendations for future research directions. Six promising research directions are identified for further investigation of LLMs for SHDG: evaluation metrics, LLM adoption, data efficiency, generalization, health equity, and regulatory challenges

Daniel Smolyak, Department of Computer Science, University of Maryland
Margret  V. Bjarnadottir,  Robert H. Smith School of Business, University of Maryland
Kenyon Crowley, Accenture Federal Services
Ritu Agarwal, Center for Digital Health and Artificial Intelligence, Carey Business School

Local Indicators of Spatial Association (LISA) analysis is a useful tool for analyzing and extracting meaningful insights from geographic data. It provides informative statistical analysis that highlights areas of high and low activity. However, LISA analysis methods may not be appropriate for zero-heavy data, as without the correct mathematical context the meaning of the patterns identified by the analysis may be distorted. We demonstrate these issues through statistical analysis and provide the appropriate context for interpreting LISA results for zero-heavy data. We then propose an improved LISA analysis method for spatial data with a majority of zero values. This work constitutes a possible path to a more appropriate understanding of the underlying spatial relationships. Applying our proposed methodology to crack cocaine seizure data in the U.S., we show how our improved methods identify different spatial patterns, which in our context could lead to different real-world law enforcement strategies. As LISA analysis is a popular statistical approach that supports policy analysis and design, and as zero-heavy data is common in these scenarios, we provide a framework that is tailored to zero-heavy contexts, improving interpretations and providing finer categorization of observed data, ultimately leading to better decisions in multiple fields where spatial data is foundational.

Eunseong Jang, The Robert H. Smith School of Business, University of Maryland
Margret Bjarnadottir, The Robert H. Smith School of Business, University of Maryland
Marcus Boyd, National Consortium for the Study of Terrorism and Responses to Terrorism, University of Maryland
S. Raghavan, The Robert H. Smith School of Business & Institute for Systems Research, University of Maryland