UCSF

Healthcare data has never been so accessible to patients and physicians, from smartphones and other remote monitoring devices to improved access for patients to their own Electronic Medical Record (EMR) history and clinical notes. Despite the ubiquity of healthcare data collection and distribution, there remains a significant gap in understanding the impacts of this data on clinical care. Insights from these digital health tools and downstream clinical decision-making processes are often only captured in medical notes, which are complex, sparse, unstructured, and difficult to model even with traditional deep learning methods. Only recently have large language models (LLMs) emerged that are capable of zero- or few-shot clinical language, without the need for large, manually annotated datasets. In this dissertation, I develop methods to apply LLMs to healthcare data, particularly for identifying points of actionable insights for both digital and pharmaceutical therapeutics. These approaches demonstrate the ways in which digital health products and passive monitoring can impact clinical care, and identify reasons for medication class switching that take into account the complexities of patient care beyond lab values and diagnosis codes. While careful, rigorous research is needed to ensure that these approaches are effective in facilitating patient care improvements and to reduce any potential for harm, the rapid pace of language model development provides an extraordinary opportunity to transform clinical practice. These new methods allow us to take an unprecedented look at the conversations, decisions, and medical expertise captured in billions of clinical notes and other clinical text, and to learn from this shared knowledge to accelerate medical research, improve clinical guidelines, and personalize patient care.