Linear methods for classification

Linear classification methods are highly prevalent in clinical neuroimaging and have been used to predict diagnosis and outcome in many brain disorders. Here, we provide a concise introduction to these methods aimed at the beginning practitioner. We introduce the two main variants: penalized linear models and probabilistic classification models, highlighting their relative strengths and weaknesses. We describe discriminative mapping, which is the ability to visualize the model coefficients in the input space and is a crucial benefit of linear models because it helps to understand which features of the data drive the predictions. We also introduce the notion of sparsity, which further assists interpretation in that it can be used to restrict the discriminative pattern to a small number of brain regions. Finally, we provide an overview of studies using linear models along with two illustrative applications using linear models to discriminate patients with autism and schizophrenia from healthy participants.