Concordance index loop optimization

Hello, I'm implementing de concordance index measure (<http://biomet.oxfordjournals.org/content/92/4/965.abstract>) to compare a regression model with tru labels, and I attempted to do a loop based implementation that is currently very slow (see below). Does anybody know a good implementation of this measure, or how to optimize this code?

Thanks in advance

CODE:

function [ci] = acan_concordance_index(targets, predicts)
  [~,i] = sort(targets,'descend');
  predicts = predicts(i);
  n = length(targets);
  total  = 0;
  norm_z = 0;
  for j=1:n
      for k=j:n
          if j ~= k
              h = step_function(predicts(j) - predicts(k));
              total = total + h;
              norm_z = norm_z + 1;
          end
      end
  end
  ci = total / norm_z;

end

 function h = step_function(diff)
    if diff > 0
        h = 1;
    elseif diff == 0
        h = 0.5;
    else
        h = 0;
    end
 end

[Geoff Hayes]

In the above code, the outer for loop is iterating over j and the inner for loop is iterating over k. Yet h is being calculated as

 h = step_function(predicts(i) - predicts(j));

Do you mean to be using i which is a vector returned from the sort function?

[Andre]

Yes Geoff, I misstyped it, and already edited. I also changed the inner loop to begin with j, as the list is already ordered. I found a nice definition of the c-index here: https://stat.ethz.ch/pipermail/r-help/2008-December/182565.html The code seems to match mine, but it is also non optimized.

Thanks for the reply.

Answer 1

[Julio Chirinos]

does anybody have a function to compute the Harrel c index in Matlab? I have reached out to the Mathworks to no avail. Any help would be greatly appreciated. Alternatively, in the function above, what is targets and predicts? how do I go from the output of coxphfit to these 2 arguments in order to use the function? Thank you!