stanfordmlgroup / chexpert-labeler

CheXpert NLP tool to extract observations from radiology reports.
MIT License
328 stars 78 forks source link

Probabilities #51

Open danielemolino opened 4 days ago

danielemolino commented 4 days ago

Hi, I was thinking if it is somehow possible to extract the probabilities of the classification, to then evaluate the AUROC

jirvin16 commented 4 days ago

This is a rule-based labeler so it does not support outputting classification probabilities.

danielemolino commented 4 days ago

Thanks! It is exactly what I wanted to hear. Because I read some works where someone was computing AUROC on the prediction obtained by the labeler, and I was trying to understand how was that possible.

Do you confirm me that this doesn't make much sense, and it is like evaluating the accuracy of the model?

jirvin16 commented 4 days ago

Not sure what works you're referring to, but this work for example does produce classification probabilities!