Mean squared error (MSE) is a measure of prediction accuracy for a particular K. It is calculated as implied by its name: take the average of the squared differences between true label and prediction. Note that MSE should be computed for out-of-sample error when determining the best K, i.e. using test rather than training data.