It's been a while since I looked at this, however my recollection is that to identify the cut-off criteria, the ROC curve is plotted (I used SPSS), and that for each coordinate of sensitivity versus (1- specificity) you also get the corresponding test result variable result (in my case it was change in symptom score), then derive each Youden's Index (multiple values for each coordinate; I also used Excel after copying and pasting SPSS output into Excel). The stress level will be on a continuum and you are after a cut-off criteria which is determined by where the ROC curve demonstrates furthest distance from the line of identity (and highest Youden's Index).

In my case it was simply reading off the test variable result (i.e. minimally important change score) which concorded with the highest Youden's Index, so I didn't need to resort to logistic regression. Not sure if that helps, but it doesn't sound like you have your cortisol values for each of your sensitivity and (1-spec) coordinates? 

Good luck

Have read on some stats forums that you have to solve it using logistic regression. But what you said makes sense - I have an additional unlabelled column in my outputs that corresponds with the same min and max range in my raw data.  So if it is standard practise to read off the corresponding concentration that relates to the youdens index I'll do that, thanks!

In my case I didn't have problems with the distribution, and there were no other test results to consider in improving the diagnostic accuracy. One next step I thought to do was to test the theoretical cut-off in new population, or to think about the sort of regression analysis you are posing...I've not done this so I will be really interested to see if anyone else can shed light on approaches here. Sorry not to be of more help