A framework has been designed to estimate the performance shift of an AI model when evaluated on a new unlabelled target domain.
We introduce a regression-based method that proposes two different domain-shift detection metrics and employs them to predict the performance shift of a model on a target domain on which ground truth labels are not available. The main advantage of this approach is that it has been designed to deal with unlabelled target data as often is in real-world clinical context. For this reason, it could address the issue of the lag between actual performance drift and labelled data becoming available.
More details can be found at: https://drive.google.com/drive/u/0/folders/1XylAlEKPIj7YX-cnBDyE6z7nmyuVgMC2
Have a look at the'main.Rmd' file!