Fig. 3.

Modelling clinical outcome prediction with low and high-dimensional methods. Mean ROC curves(solid lines) and 95% confidence intervals (dotted lines) for predicting the disease label assuming a model of anatomical dependence involving damage to either BA39 or BA44 (left panel), or BA37 or BA38 (right panel), derived from the data reported in Mah et al. (2014). In the mass-univariate case (in black), the predictions were generated by applying a voxel-wise Fisher exact test to randomly sampled subsets of 407 lesions, thresholded at p < 0.001 uncorrected, and then testing on the remaining 174, generating a single predictive value for each test lesion using Fisher's method. In the high-dimensional multivariate case (in red), the predictions were generated by estimating the weighting of each voxel with a linear support vector machine (cost parameter= − 14), and applying the weights to each test image. The mean curves and their confidence intervals are derived from 200 random iterations across the full 581 lesion dataset. Note that the multivariate approach is markedly superior. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)