Learning performances with different algorithms and different implementations. All bars represent results of a repeated (ten times) ten fold cross validation. A: Categorical learning with different learning algorithms without feature selection. The V50 values were divided into seven classes based on their values. The learning was done without feature selection. B: Improvement of KNN prediction accuracies in different implementations. Results of KNN classification without feature selection, with the filter algorithm, with the wrapper algorithm, and with outlier selection in combination with the wrapper algorithm are shown. Both feature selection algorithms improved learning performance. The best learning accuracy was obtained using the KNN classifier combined with wrapper and removal of four outlier sequences. It yields a mean absolute errors of 7.0 mV with the new dataset (Dataset 2) of 54 VKC sequences.