ABSTRACT

Sudden Unexpected Death in Epilepsy (SUDEP) is a significant health concern in epilepsy. In the United States, it's estimated that about 3.4 million people, covering both kids and adults, live with epilepsy and continually experience seizures. With a mortality rate of approximately 1 in every 1,000 epilepsy patients annually, SUDEP's unpredictability and severity necessitate a deeper understanding of its underlying mechanisms. The American Academy of Neurology published SUDEP guidance to highlight the importance of communicating the risk of SUDEP to patients with epilepsy, since risks may be modifiable if epilepsy is properly managed and treated. However, there is a lack of individualized, evidence-based tools for systematic SUDEP risk assessment.

This dissertation develops a set of integrated informatics approaches to enable personalized SUDEP risk assessment using the Center for SUDEP Research (CSR) clinical data resource. Large language model (LLM)-based natural language processing (NLP) methods are developed to extract clinically relevant SUDEP risk markers from unstructured Epilepsy Monitoring Unit (EMU) evaluation reports, including seizure frequency, seizure types, and developmental intellectual disability, while also performing ontology-based normalization to standardize extracted concepts into structured clinical representations. Building on these extracted variables, an automated framework is introduced to generate individualized SUDEP risk marker reports based on the SUDEP-7 inventory, providing a structured representation and scoring of patient-specific risk profiles. Finally, machine learning-based predictive models are leveraged for SUDEP risk assessment by integrating these extracted risk markers with additional CSR clinical data elements.

Such systematic SUDEP risk assessment tools are expected to help clinicians and patients better communicate and manage potentially modifiable risks, ultimately leading to overall reduced SUDEP mortality and improved epilepsy patient care.