Performing Named Entity Recognition on the i2b2 2010 dataset using clinicalBERT and an anchoring approach, warm-starting clinicalBERT embeddings with noisy-labeled MIMIC III notes using UMLS terms.
A way of performing stable regression based on a new risk measure, the Entropic Value at Risk (EVaR), instead of the Conditional Value at Risk (CVaR) used in (Bertsimas and Paskov, 2019).
An extension of Stable Regression from (Bertsimas and Paskov, 2019) to SVMs and Logistic Regression, allowing coefficients to be up to 200% more stable (in terms of standard deviation). Implemented in Python and Julia.