# Method to determine statistical significance between two machine learning methods

I want to compare the accuracy of two machine learning (ML) methods, each of which have two classifiers within them. I have six subjects to test these two ML methods against. The same subjects, data, and everything in between is the same for both ML methods.

I've run the tests and the results are presented in this form, where each subject/classifier row is a percentage of accuracy:

ML method 1            ML method 2
-------------------    -------------------
sub 1, classifier 1    sub 1, classifier 1
sub 2, classifier 1    sub 2, classifier 1
...                    ...
sub 6, classifier 1    sub 6, classifier 1
sub 1, classifier 2    sub 1, classifier 2
sub 2, classifier 2    sub 2, classifier 2
...                    ...
sub 6, classifier 2    sub 6, classifier 2


To get the results subject-wise, I take the average of each classifier per subject (mean(subx_class1, subx_class2), where x is subject 1 to 6), which gives me:

ML method 1            ML method 2
-------------------    -------------------
sub 1                  sub 1
sub 2                  sub 2
sub 3                  sub 3
sub 4                  sub 4
sub 5                  sub 5
sub 6                  sub 6


My questions is, what would be the best statistical method to use in this example to determine if the two ML methods produce different accuracies which are statistically significant? I don't have a great deal of knowledge in this area, and I've mostly dealt with paired t-tests before, which I don't think I can do in this example due to the data not being subject specific.

Thanks very much for any help.