NLM DIR Seminar Schedule

Abstract:

Accurate prediction systems are highly sought after in clinical care due to their potential benefits, such as aiding clinicians in decision-making, reducing costs, and enabling personalized treatment. However, obtaining annotated data for clinical risk prediction, particularly temporal information, is often expensive and time-consuming. In this paper, we focus on leveraging large language models (LLMs) to extract case reports and generate clinical events with associated timestamps from clinical texts, thereby reducing the costs of patient history collection and annotation. Specifically, we utilize GPT-4 to simulate a physician reading clinical texts and generating clinical event-timestamp pairs. These generated pairs are then used to perform a two-stage fine-tuning of LLaMA 2 on both generated and real data for heart failure prediction. Experimental results on a large EHR dataset, comprising 14 million visits by 263,000 patients over an 8-year period, demonstrate that our model, trained with generated clinical event-timestamp pairs, outperforms the baseline model trained solely on real data. Moreover, our approach is adaptable to other clinical risk prediction tasks in real-world settings.