Maskininlärning för diagnosticering av perifer neuropati

Abstract

This report investigates the possibility of diagnosing peripheral neuropathy with the help of non-parametic classification methods. Peripheral neuropathy is a disease state characterized by damage on the nerves furthest out in the nervous system, with symptoms first occuring in the feet. The data used in this project comes from Dr. William Kennedys research group at University of Minnesota. The data contains 401 observations of 120 healthy controls and 65 individuals with presumed peripheral neuropathy due to chemotherapy, (where 18 individuals have been confirmed having peripheral neuropathy through other examination procedures). The data is collected with a dynamic sweat test, a new diagnostic method to discover unusual sweating patterns and therefore also peripheral neuropathy. In this project we compare three different machine learning methods to classify subjects as sick (peripheral neuropathy) and healthy (no peripheral neuropathy): k-NN, random forest and neural networks. These methods differ in their complexity, all with their disadvantages and advantages. To evauluate which classification method that works the best a cross-validation was performed, with a modified version of Cohen’s kappa. How good these classification methods perform depends on which measuring area the data comes from, either foot, calf or foot and calf combined. The best classification method was shown to be random forest, this for the calf-measurements where the covarariates are chosen by backward stepwise selection. This method correctly classifies 67% of the sick individuals and 96% of the healthy controls. With the best model trained on foot-measurements most undetermined sick individuals are being classified as sick, while for the best model trained on calf-measurement most of the undetermined sick individuals are classified as healthy. This could hint towards that the symptoms of peripheral neuropathy first appears in the feet, something that is in line with the clinical reality.