Clinically Grounded Privacy Evaluation of Medical LMs
Medical language models can leak sensitive patient information, but standard memorization metrics may not reflect realistic privacy risks. We introduce a clinically grounded framework that evaluates leakage across levels of adversarial access, revealing substantial patient-specific and sensitive-diagnosis leakage while showing that exact-match memorization can overstate disclosure due to templated clinical text.
Sep 9, 2026
MINTY: Rule-based Models that Minimize the Need for Imputing Features with Missing Values
We propose MINTY, a rule-based model that learns to avoid reliance on missing features by using disjunctive rules as replacements, maintaining interpretability and predictive performance while reducing dependence on imputation.
May 3, 2024
Sharing Pattern Submodels for Prediction with Missing Values
Sharing Pattern Submodels (SPSM), a method that balances pattern specialization and information sharing by enforcing parameter sharing through sparsity-inducing regularization, improving robustness to missing values while maintaining interpretability and predictive power.
Jun 26, 2023
Learning replacement variables in interpretable rule-based models
MINTY is a method for learning compact and interpretable rule models that handle missing values at both training and test time by using replacement variables, enabling robust predictions and balancing interpretability with predictive performance, particularly in clinical applications.
Jun 20, 2023