Human brain lesion-deficit inference remapped

Abstract

Our knowledge of the anatomical organization of the human brain in health and disease draws heavily on the study of patients with focal brain lesions. Historically the first method of mapping brain function, it is still potentially the most powerful, establishing the necessity of any putative neural substrate for a given function or deficit. Great inferential power, however, carries a crucial vulnerability: without stronger alternatives any consistent error cannot be easily detected. A hitherto unexamined source of such error is the structure of the high-dimensional distribution of patterns of focal damage, especially in ischaemic injury-the commonest aetiology in lesion-deficit studies-where the anatomy is naturally shaped by the architecture of the vascular tree. This distribution is so complex that analysis of lesion data sets of conventional size cannot illuminate its structure, leaving us in the dark about the presence or absence of such error. To examine this crucial question we assembled the largest known set of focal brain lesions (n = 581), derived from unselected patients with acute ischaemic injury (mean age = 62.3 years, standard deviation = 17.8, male:female ratio = 0.547), visualized with diffusion-weighted magnetic resonance imaging, and processed with validated automated lesion segmentation routines. High-dimensional analysis of this data revealed a hidden bias within the multivariate patterns of damage that will consistently distort lesion-deficit maps, displacing inferred critical regions from their true locations, in a manner opaque to replication. Quantifying the size of this mislocalization demonstrates that past lesion-deficit relationships estimated with conventional inferential methodology are likely to be significantly displaced, by a magnitude dependent on the unknown underlying lesion-deficit relationship itself. Past studies therefore cannot be retrospectively corrected, except by new knowledge that would render them redundant. Positively, we show that novel machine learning techniques employing high-dimensional inference can nonetheless accurately converge on the true locus. We conclude that current inferences about human brain function and deficits based on lesion mapping must be re-evaluated with methodology that adequately captures the high-dimensional structure of lesion data.

Keywords: focal brain injury; ischaemic brain injury; lesion-deficit inference.

Figure 1

Illustration of how stereotyped patterns of brain damage (schematized in grey) across a set of patients can hypothetically mislocalize damage of any part of critical area A (in dotted lines) to the non-critical area B (in dotted lines). This will happen whenever the spatial variability of damage to a non-critical area is less for the group or factor of interest than for the critical area. Such stereotypy of damage—a hidden deep structure in the data—may occur where the lesions follow a consistent non-neural architecture, as is the case with vascular lesions.

Figure 2

Three-dimensional vector plot of the direction (colour map) and magnitude (length of arrow) of mislocalization at adequately sampled voxels within three representative planes (left axial, top coronal, bottom sagittal), based on a sample of 581 acute stroke lesions, normalized into standard stereotactic space and mirrored onto one hemisphere (see ‘Materials and methods’ section for details). The value at each voxel was calculated by labelling the stack of 581 lesioned volumes as being ‘affected’ or ‘unaffected’ depending on whether or not that voxel fell within the lesion in each volume, running a standard voxel-wise Fisher’s exact test-based mass-univariate analysis on the two groups, and identifying the centre of mass of the resultant significant cluster, identified by the asymptotic P-value thresholded at a Bonferroni corrected P < 0.01. This procedure was performed at all voxels hit more than three times in the data set. Each arrow points from the true location of a voxel in the brain to the location where the mass-univariate model erroneously places it. The colour map corresponds to the orientation of this error vector in the visualized plane. Note that the mislocalization tends to follow the organization of the vascular tree, with clusters corresponding to the branches of the middle cerebral, anterior cerebral, and posterior circulations. See ‘Materials and methods’ section for details. See Supplementary material for manipulable 3D versions of these images.

Figure 3

(A) Three-dimensional plots of the voxels identified as significantly associated with a hypothetical deficit—given damage to either BA 39 or BA 44 at ≥20% of the volume of either—by a voxel-wise mass-univariate analysis of the sample of 581 acute stroke lesions (red cubic glyphs). As before, Fisher’s exact test was used, thresholded at a level such that the volume of surviving voxels equalled 20% of the volume of BA 39 and BA 44 (each area is shown as a black wireframe). Note that the centre of mass of the significantly associated region falls in neither Brodmann area, but in the region of the superior temporal gyrus (STG, grey wireframe). See ‘Materials and methods’ section for details. In grey is an outline of an axial slice traversing BA 44 and BA 39, shown here purely to give an indication of the relative position of the two areas in the axial plane. (B) Three-dimensional plots of the voxels identified as heavily weighted in the classification process—given damage to either BA 39 or BA 44 at ≥20% of their total volume—by a high-dimensional multivariate analysis of the sample of 581 acute stroke lesions based on a linear support vector machine (blue cubic glyphs). The voxels shown are thresholded so as to yield the same number of surviving voxels as in the mass univariate analysis depicted in A. Note that the mislocalization observed with the mass-univariate approach is no longer seen. See ‘Materials and methods’ section for details. See Supplementary material for manipulable 3D versions of these images.

Figure 4

(A) Three-dimensional plots of the voxels identified as significantly associated with a hypothetical deficit—given damage to either BA 37 or BA 38 at ≥20% of the volume of either—by a voxel-wise mass-univariate analysis of the sample of 581 acute stroke lesions (red cubic glyphs). As before, Fisher’s exact test was used, thresholded at a level such that the volume of surviving voxels equalled 20% of the volume of BA 37 and BA 38 (each area is shown as a black wireframe). Note that the centre of mass of the significantly associated region falls in neither Brodmann area. See ‘Materials and methods’ for details. In grey is an outline of an axial slice as in Fig. 3, shown here purely to give an indication of the relative position of the two areas in the axial plane. (B) Three-dimensional plots of the voxels identified as heavily weighted in the classification process—given damage to either BA 37 or BA 38 at ≥20% of their total volume—by a high–dimensional multivariate analysis of the sample of 581 acute stroke lesions based on a linear support vector machine (blue cubic glyphs). The voxels shown are thresholded so as to yield the same number of surviving voxels as in the mass univariate analysis depicted in A. Note that the mislocalization observed with the mass-univariate approach is much less pronounced. See ‘Materials and methods’ section for details. See Supplementary material for manipulable 3D versions of these images.