Kohonen-Based Topological Clustering as an Amplifier for Multi-Class Classification for Parkinson’s Disease

Classifying the degree of Parkinson's disease is an important clinical necessity. Nonetheless, current methodology requires manual (and subjective) evaluation by a trained clinical expert. Recently, Machine Learning tools have been developed that can produce a classification of the presence of PD directly from the speech signal in an automated and objective fashion. However, these methods were not sufficient for the classification of the degree of the disease. In this work, we show how to apply and leverage topological information on the both the label space and the feature space of the speech signal in order to solve this problem. We address the problem by performing topological clustering (using a version of the Kohonen Self Organizing Map algorithm) of the feature space and then optimizing separate multi-class classifiers on each cluster. Using these methods, we can reliably train our system to classify new speech signal data to more than the 70% level on a 7 degree classification (where random level is 14%) which is close to the obtainable accuracy on the simple 2 class classification.