Research Projects

Project theme: Understanding how large language models reason over time in complex, real-world electronic health records (EHRs). Project Motivation Clinical care unfolds over days to weeks, not single notes. Diagnoses emerge, treatments change, and patient status evolves.

This lab manual outlines the values, structure, and research practices that guide work in the LARK Lab. It reflects our problem-driven approach to clinical NLP and AI research, our emphasis on reproducibility and transparency, and our commitment to building tools that operate within real healthcare systems.

In this work, we delve into the Chain-of-Thought (CoT) Distillation problem, presenting it through the lens of information theory. Our approach re-examines CoT distillation and places it within an information bottleneck framework.

Large Language Models (LLMs) like GPT-4 have demonstrated significant advancements in tasks such as Multiple Choice Question (MCQ) Answering, including those within medical domains like MedQA. However, our recent research reveals a positional bias in earlier models such as GPT-2, where they consistently select the first answer choice (“A”) regardless of the actual correct answer’s position.

Diagnostic errors affect 12 million patients in the U.S. and contribute to 80,000 deaths per year. The main causes for diagnostic errors include cognitive biases introduced by healthcare providers, miscommunication between healthcare teams, lack of access to key data, and not recognizing time-sensitive data in the electronic health record (EHR).