Reproduce “An Investigation of Memorization Risk in Healthcare Foundation Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sana Tonekaboni, Lena Stempfle, Adibvafa Fallahpour, Walter Gerych, Marzyeh Ghassemi

Researchers from MIT, the Broad Institute, and the Vector Institute developed a six-test black-box framework to assess memorization risks in Electronic Health Record Foundation Models (EHR-FMs), revealing that while a benchmark model showed low embedding-level leakage, generative memorization of sensitive attributes increased with prompt detail, particularly for specific rare conditions or when predictions relied on patient-specific factors.

Key insights

  • Defining and evaluating memorization in healthcare AI requires a nuanced approach, distinguishing between benign population-level generalization and privacy-compromising patient-level recall.
  • The extent of memorization risk in EHR-FMs is highly dependent on the specificity and volume of input information provided, with more context leading to higher generative leakage potential.
  • Perturbation tests and sub-population analyses are crucial for identifying if sensitive information leakage stems from specific individual memorization or broader statistical patterns, highlighting vulnerable patient groups.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.