McNemar's test seems widely misunderstood and I would argue misapplied. It is useful for method comparison, but  it is not checking accuracy. Accuracy (correct classification rate) is the appropriate statistic for this, then sensitivity.specificity, PPV, NPV, PDLR, NDLR are then used to examine various angles of positive and negative results depending on the application's needs. To me attempting to use McNemar's test for accuracy comparison is at best misguided, at worst p hacking (one knows traditional comparison of accuracy won't be significant so they try a 'soft' method to get a significant p value). I return to this to justify my statements at the end, but onto what McNemar's test does:

McNemars test is checking if the distribution of disagreements between two (or more for higher dimensional equivalents of the test) methods is skewed. I.e. are the disagreements imbalanced between the tests?

It ignores the agreement diagonal of the confusion matrix. As such it is especially useful (but not limited to) in comparing high accuracy methods as it ignores the agreements (which swamp disagreements in high accuracy tests), you see from the formula that f11 and f22 are not included nor the total number of results.

If you have multiple comparisons you should be using the full 6x6 confusion matrix rather than a sequence of 2x2 confusion matrices. This is then used with the Cochran's Q test rather than McNemar's. If sticking with McNemars then some false positive rate correction is needed for multiple comparisons.

f12 represents the number of samples classified by method A as belonging to class 1 that are classified as class 2 by method B - completely regardless of whether true or not. f21 is the number of samples where method A gives class 2 while method B gives class 1, again irrespective of truth or not.

It is possible to use McNemars vs ground truth (where either method A or B is ground truth), but you need clear pre-hoc justification for doing this as it is not the most natural choice. Even so, in this case it is not measuring 'accuracy' and if method A is ground truth and 2 is positive then f12 becomes false positives, f21 becomes false negatives. In this case it is measuring how skewed the test is between false positives and false negatives and you would need to work out the theoretical implications of this before applying it. Again, the various traditional positive/negative summary stats are much more interpretable for handling false positive and false negatives and therefore more useful.

McNemar's test seems widely misunderstood and I would argue misapplied. It is useful for method comparison, but  it is not checking accuracy. Accuracy (correct classification rate) is the appropriate statistic for this, then sensitivity.specificity, PPV, NPV, PDLR, NDLR are then used to examine various angles of positive and negative results depending on the application's needs. To me attempting to use McNemar's test for accuracy comparison is at best misguided, at worst p hacking (one knows traditional comparison of accuracy won't be significant so they try a 'soft' method to get a significant p value). I return to this to justify my statements at the end, but onto what McNemar's test does:

McNemars test is checking if the distribution of disagreements between two (or more for higher dimensional equivalents of the test) methods is skewed. I.e. are the disagreements imbalanced between the tests?

It ignores the agreement diagonal of the confusion matrix. As such it is especially useful (but not limited to) in comparing high accuracy methods as it ignores the agreements (which swamp disagreements in high accuracy tests), you see from the formula that f11 and f22 are not included nor the total number of results.

If you have multiple comparisons you should be using the full 6x6 confusion matrix rather than a sequence of 2x2 confusion matrices. This is then used with the Cochran's Q test rather than McNemar's. If sticking with McNemars then some false positive rate correction is needed for multiple comparisons.

f12 represents the number of samples classified by method A as belonging to class 1 that are classified as class 2 by method B - completely regardless of whether true or not. f21 is the number of samples where method A gives class 2 while method B gives class 1, again irrespective of truth or not.

It is possible to use McNemars vs ground truth (where either method A or B is ground truth), but you need clear pre-hoc justification for doing this as it is not the most natural choice. Even so, in this case it is not measuring 'accuracy' and if method A is ground truth and 2 is positive then f12 becomes false positives, f21 becomes false negatives. In this case it is measuring how skewed the test is between false positives and false negatives and you would need to work out the theoretical implications of this before applying it. Again, the various traditional positive/negative summary stats are much more interpretable for handling false positive and false negatives and therefore more useful.

.

you might find the following tutorial useful :

http://www.icmla-conference.org/icmla10/CFP_Tutorial_files/jose.pdf

(although it does not mention Mc Nemmar test ... but James explained above why there is indeed no reason to mention it !)

.

You may find this paper useful. The authors used the same test.

A wavelet optimisation approach for ECG signal classification (2012) by A. Daamouche et al

As mentioned by Beattie, the test is used to compare two classifiers and check whether classifier A outperforms classifier B significantly.