Task-related principal component analysis: formalism and illustration.

We address the application of a modified form of Principal Component Analysis (PCA) to data which is characterized by sparse, but known, event times. The sparsity of such event times makes it unlikely that they have a major contribution to the overall covariance in the data, and standard PCA components generated from this covariance may not give us useful insight about the task. A simple method is shown here which generates an orthogonal, "task-related PCA" (trPCA) transform based upon correlations between non-simultaneous, event-time locked, subsets of data. Non-simultaneity is the constraint that epochs of data are only compared to epochs of data from other points in time, which explicitly selects for reproducible effects. The prescription for trPCA is presented within the context of a fusiform face area experiment for illustration. In this experiment, a reproducible, face-stimulus specific, negative potential deflection is observed 200ms (N200) after presentation. We demonstrate how this N200 phenomenon, initially distributed across a subtemporal electrocorticographic (ECoG) array, may be isolated in a single component using trPCA.