Support‎ > ‎

Paired SVM

We are pleased to announce a new Methods Core release: the Paired Support Vector Machine Toolbox. Support vector machine (SVM) is a powerful multivariate method for classifying imaging data (or any data) into groups, e.g., Depressed versus Healthy, Responders versus Non-Responders, and so on.

Up until now, however, there has never been a good method to apply SVM to within-subject designs. Daniel Kessler and I have developed an approach called Paired SVM that addresses this problem. In the same way that a paired t-test "subtracts out" the between-subject variance, Paired SVM performs a similar trick. Here are the kinds of problems that Paired SVM can address: 
  1. You scanned the same subject on and off a medication, and you now want to know the multivariate neurosignature of the drug.
  2. You scanned the same subject pre and post some intervention (or just at two points in time) and you want to know the 'signature of change' between the two time points.
  3. You observed the same subject in the task versus control condition of an imaging paradigm. You want to characterize the distributed neurosignature produced by task.

Paired SVM has now been placed in your lab's Methods Core repository. If you are interested in applying Paired SVM to your data, send me an email. This is a so-called "Advanced Method" partially developed with Methods Core resources, so in accordance with the Use Agreement for these tools, we request that you involve Daniel Kessler and Chandra Sripada as collaborators. 

Once you have an accurate classifier, the next step is to visualize it in some way to see on what basis it is making its "decisions". We have also developed a number of visualization schemes for the outputs from Paired SVM (these apply equally well to Unpaired SVM as well). So let us know if you are interested in that.

Validation Information
In a recent NeuroImage paper (available here), we describe the Paired SVM method in detail and apply it to identify the network neurosignatures of acute administration of methylphenidate. We have also conducted a large simulation that validates the method. We show that for within-subject data, Paired SVM is always as good and usually much better than Unpaired SVM across a wide range of conditions. We are submitting this simulation result for publication soon, but the main figure is available here.