Reverse Engineering FDA Decision-Making Using Large Language
Models
Name
kwag-kwags-sm-orc-2026-thesis.pdf
Size
782.65 KB
Format
Adobe PDF
Checksum (MD5)
547a3cef1b6373d6ffe8710f8255a5b7
Author(s)
Kwag, Shauna Seoyeong
Advisor(s)
Lo, Andrew
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
All new drugs in the United States must undergo a rigorous approval process conducted by the U.S. Food and Drug Administration (FDA). In evaluating whether to approve a new therapy, regulators must weigh evidence on efficacy and safety while facing an inherent trade-off between Type I errors (approving an ineffective or unsafe drug) and Type II errors (rejecting an effective one). Because the FDA does not explicitly disclose the rationale underlying its decisions, approvals and rejections can at times appear controversial or opaque. This thesis adopts a quantitative, reverse-engineering approach to infer the implicit trade-offs reflected in FDA decision-making. We focus on three key factors commonly considered in regulatory evaluation: drug efficacy, safety, and therapeutic context. To study these factors, we analyze advisory committee meetings in which independent experts review clinical evidence and provide non-binding recommendations to the FDA. Using generative AI, we systematically parse meeting transcripts and encode committee member discussions into numerical sentiment scores for efficacy and safety. We then combine these measures with FDA approval outcomes in logistic regression and Bayesian hierarchical models to examine how regulatory decisions relate to advisory committee perspectives. Our results suggest that both therapeutic classification and disease severity influence whether efficacy or safety is prioritized in regulatory decisions, and that decision-making patterns vary across therapeutic areas. These findings provide a quantitative framework for understanding the implicit balance between Type I and Type II errors in regulatory decision-making and suggest that more explicit decision-theoretic frameworks could improve transparency in how drug approval decisions are made.
MIT Department
Massachusetts Institute of Technology. Operations Research Center
Sloan School of Management
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
