I haven't reviewed the paper, so I am not qualified to be overly critical. I am concerned that we are trapped in the "silver bullet" approach to a very difficult problem. Classification is best with several variables, but we aren't given such a choice. Here, there were three tests, and the weakest was withdrawn from consideration.

If there was robust introduction of variables, some being scaled observations (STE 4, STD 3, TWI 2, indeterminate 1, NSR 0), previous MI (0,1), stabbing (0,1), etc. you could classify the data using only a third of the patients. The classification will result in clusters, which will aggregate based on distances of the centroids in 3-D space. Then you have the optimum number of classes, which is valued by Bayes Information Criteria and by Akaike's Information Criteria. These have been predicated based on the work of Solomon Kullback. You can then take each group and look at the proximity to the diagnoses used for the ROC curve. Once that is settled, you should have a predictive model that has combinations of variables with higher or lower weights. This is still only based on part of the database. You can now use the remainder of the data to prove the validity, and to obtain probabilities for the diagnoses.