CI of AUC with transformed specificity

[seanso95]

Is your feature request related to a problem? Please describe.
Hi, I have been working with various ROC curves to train some machine learning models. I like the efficiency of the ci() functions implemented in pROC. However, I have been using a square root transformed specificity to emphasise early detection in my ROC curves.

I can update the rocobject$sensitivies to the preferred scale, and recalculate AUC with auc(), but cannot use ci.auc() as it either uses the embedded AUC (linear scale) or if reuse.auc=FALSE is set, recalculates AUC using the original predictors/labels as far as I can tell.

Describe the solution you'd like
I would love if we could update the roc object for ci() calculations!

Describe alternatives you've considered
Alternatively, allow passing in some function as an argument to roc(), similar to the trans argument in some of ggplot's axis functions?

Thank you for any help towards this feature!
Sean

[xrobin]

That's a tough one.

My gut feeling is that square-root specificity / sensitivity should go to the coords function as coordinates.

But it wouldn't give you an AUC. So we would need auc.coords function and a corresponding ci function.

Doable but it's a long shot. I'll keep it open as this would be nice to have but don't expect to see anything soon.