Data. Can't make many suggestions without actual data.
We can calculate a mean Y value for multiple values of X. But if there are only 4 points at a given X then a confidence interval may be very wide if given only the curve values.
Probably a better approach would be to combine the data used to create all the ROC and then create the ROC for that combined data.
By c-statistic do you mean the c concordance index?
What do you mean by a "reference group"? Do you have yet another ROC for that group? Do you mean to display that curve along with the "mean ROC" requested? Likely possible but again data. I think we would be combining data in some form to send to SGPLOT requesting multiple curves.