Clinically Grounded Privacy Evaluation of Medical LMs

Medical language models (LMs) can memorize and reproduce protected health information, but privacy evaluations often focus on recovery of training text rather than disclosure under realistic threat models. We introduce a clinically grounded framework that evaluates leakage along a graded axis of adversarial access, ranging from publicly inferable demographics to leaked note fragments. At each tier, we measure verbatim memorization of patient-specific text and semantic leakage of sensitive diagnoses. Applying the framework to an LM pretrained on 378k clinical notes, we find that routine encounter metadata (i.e. name, date of birth, provider, practice, visit date) elicits high rates of verbatim memorization across a patient's timeline and sensitive-diagnosis recovery (AUROC 0.91 for abortion, 0.81 for HIV). At the same time, exact-match memorization can overstate disclosure: 36% of memorized tokens reflect templated documentation. Our work highlights the risks of training on longitudinal clinical data, providing a practical framework for contextual privacy evaluation of medical LMs.

Blueprinting the Future: Automatic Item Categorization using Hierarchical Zero-Shot and Few-Shot Classifiers

In testing industry, precise item categorization is pivotal to align exam questions with the designated content domains outlined in the assessment blueprint. Traditional methods either entail manual classification, which is laborious and error-prone, or utilize machine learning requiring extensive training data, often leading to model underfit or overfit issues. This study unveils a novel approach employing the zero-shot and few-shot Generative Pretrained Transformer (GPT) classifier for hierarchical item categorization, minimizing the necessity for training data, and instead, leveraging human-like language descriptions to define categories. Through a structured python dictionary, the hierarchical nature of examination blueprints is navigated seamlessly, allowing for a tiered classification of items across multiple levels. An initial simulation with artificial data demonstrates the efficacy of this method, achieving an average accuracy of 92.91% measured by the F1 score. This method was further applied to real exam items from the 2022 In-Training Examination (ITE) conducted by the American Board of Family Medicine (ABFM), reclassifying 200 items according to a newly formulated blueprint swiftly in 15 minutes, a task that traditionally could span several days among editors and physicians. This innovative approach not only drastically cuts down classification time but also ensures a consistent, principle-driven categorization, minimizing human biases and discrepancies. The ability to refine classifications by adjusting definitions adds to its robustness and sustainability.