Presentation at Machine Learning for Healthcare 2023 in New York on our work on rank-based compatibility. During the conference I presented a brief spotlight talk introducing our work and also had the chance to present a poster going into more detail. I’ve included copies of both in this blog post.
You can find a link to the post about the paper here.
A recording of the spotlight intro video.
Spotlight presentation slides
Updating clinical machine learning models is necessary to maintain performance, but may cause compatibility issues, affecting user-model interaction. Current compatibility measures have limitations, especially where models generate risk-based rankings. We propose a new rank-based compatibility measure and loss function that optimizes discriminative performance while promoting good compatibility. We applied this to a mortality risk stratification study using MIMIC data, resulting in more compatible models while maintaining performance. These techniques provide new approaches for updating risk stratification models in clinical settings.