How about this method: Figure out x # of levels for inclusion (3 for my example), #1 = say good if on a hit,#2 - should be for a hit,#3 - must have for a hit, bear with me. The highest level, give it a weight of say 60% of your score, the middle level could have say 25% and the bottom level with 15% wieght.

so % of level 3's times .60 +
% of level 2's times .25 +
% of level 1's times .10

% of is figured by # that have that critera divided by the total number of EQ's in the study.

This way you would get a number from 1-100 that
would show how accurate your prediction is.

What is nice is that as you throw in new critea, you can give it a lower "hit" value, and it won't schew the overall number too badly. Also allows you to manipulate the importance of one criterea to another, changing them and seeing the change.
As well if a fluke or two sneaks in (including data entry errors) that matches for reasons unknown, after the database starts growing, even this won't schew the result, since as the database grows, the overall impact of one episode decreases because of the way each part is calculated.

George