Soroosh Tayebi Arasteh
2 indexed papers
Publications per year
Top categories
Frequent co-authors
Research Timeline
The paper demonstrates that clinical vision-language models (VLMs) pose a significant privacy risk by allowing de-identified images to be re-linked to original reports, and proposes a targeted differential privacy (DP) fine-tuning method to mitigate this risk while preserving clinical utility.
This paper evaluates the ability of large language models to recover and express evidence grades from clinical claims, finding that while they can recover the grades, they do not consistently express them.
Papers
The strength of clinical evidence is recoverable from language model representations but not from their stated grades
This paper evaluates the ability of large language models to recover and express evidence grades from clinical claims, finding that while they can recover the grades, they do not consistently express…