Ted Pak – Massive mining of clinical notes with large language models to improve antibiotic choice for sepsis
Category: Resident
The project adapts large language models (LLMs) into a clinical informatics pipeline that can process thousands of notes in electronic medical records. The goal is to unlock completely new scales of patient data for clinical researchers to analyze in large cohort studies, where we test for associations between certain patient circumstances and outcomes in order to improve guidelines on management. The largest studies of antibiotics for sepsis have all ignored clinical notes in favor of numeric data like vitals and lab signs, because it was considered too labor-intensive for humans to read so many notes. Symptoms clustered into recognizable clinical syndromes, which correlated with each patient’s ultimate diagnosis codes. Validation of a random sample showed the model had good precision, high specificity, and high accuracy. Notably, symptoms correlated with differing risks of drug-resistant infection.
Arya Rao – The Future Patient Persona: An Interactive, Large Language Model-Augmented Harvard Clinical Training Companion
Category: Student
The project is based on an artificial intelligence (AI) technology designed to simulate patient interactions for medical students. It uses advanced language models trained on custom medical education materials to generate interactive patient scenarios for medical education. The Large Language Model allows students to practice clinical reasoning, patient interaction skills, and receive immediate and actionable feedback, all in a controlled digital environment. SP-LLMs aim to supplement traditional standardized patient encounters in medical education, offering a cost-effective and stress-reducing alternative and/or supplement to traditional objective structured clinical examinations (OSCEs).
