Some problems with construction of the k-NN classifier for recognition of an experimental respiration pathology

An objective of the work is to demonstrate some difficulties with construction of a classifier based on the k-NN rule. The standard k-NN classifier and the parallel k-NN classifier have been chosen as the two most powerful approaches. This kind of classifiers has been applied to automatic recognition of diaphragm paralysis degree. The classifier construction consists in determination of the number of nearest neighbors, selection of features and estimation of the classification quality. Three classes of muscle pathology, including the control class, and five ventilatory parameters are taken into account. The data concern a model of the diaphragm pathology in a cat. The animals were forced to breathe in three different experimental situations: air, hypercapnic and hypoxic conditions. A separate classifier is constructed for each kind of the mentioned situations. The calculation of the misclassification rate is based on the leave one out and on the testing set method. Several computational experiments are suggested for the correct feature selection, the classifier type choice and the misclassification probability estimation.