When creating a ROC curve, how do I tell Prism whether an increased or decreased test level is abnormal?

Prism figures it out. You tell it which column of data is for controls and which is for patients. It figures out whether the patients tend to have higher or lower test levels. This way Prism makes sure that the area under the ROC curve is always greater than 0.5, which is the area you'd expect with a worthless test that doesn't discriminate at all between patients and controls.

If you think Prism made the wrong choice, you can correct it by simply subtracting the area under the ROC curve (and both ends of its confidence interval) from 1.0. In this case the area will be less than 0.5, telling you that the test does worse than chance at discriminating patients from controls.