Classification of Neurodegenerative Diseases Using Gaussian Process Classification with Automatic Feature Determination

There has been a growing interest in applying pattern classification methods to discriminate patients with neurodegenerative diseases from normal controls using structural MRI data. Despite an impressive array of publications, most applications are framed as discrimination problems with categorical class decisions. For many clinical applications, probabilistic estimation would be more useful, especially when predictions from other models and measurements are combined. The Gaussian Process (GP) model is a Bayesian learning algorithm and is closely linked with classical variance component estimation. The Bayesian formulation allows the model to automatically determine model hyper-parameters by means of maximizing the “marginal likelihood” or the “evidence of the data”. We utilized such a formulation by using marginal likelihood maximization to weight the importance of different grey matter (GM) regions automatically in the training processes, without additional cross validation. The algorithms were applied to two sets of patients with Huntington´s disease and normal controls. Classification accuracy was improved from 70% to 73% with automatic feature selection for the first dataset, and was improved from 58% to 69 % for the second dataset.