Metabolic lesion-deficit mapping of human cognition

Abstract

In theory the most powerful technique for functional localization in cognitive neuroscience, lesion-deficit mapping is in practice distorted by unmodelled network disconnections and strong 'parasitic' dependencies between collaterally damaged ischaemic areas. High-dimensional multivariate modelling can overcome these defects, but only at the cost of commonly impracticable data scales. Here we develop lesion-deficit mapping with metabolic lesions-discrete areas of hypometabolism typically seen on interictal 18F-fluorodeoxyglucose PET imaging in patients with focal epilepsy-that inherently capture disconnection effects, and whose structural dependence patterns are sufficiently benign to allow the derivation of robust functional anatomical maps with modest data. In this cross-sectional study of 159 patients with widely distributed focal cortical impairments, we derive lesion-deficit maps of a broad range of psychological subdomains underlying affect and cognition. We demonstrate the potential clinical utility of the approach in guiding therapeutic resection for focal epilepsy or other neurosurgical indications by applying high-dimensional modelling to predict out-of-sample verbal IQ and depression from cortical metabolism alone.

Keywords: 18F-FDG PET imaging; depression; epilepsy; intelligence; lesion-deficit mapping.

Figure 1

Covariance structure of metabolic lesions: local dependency. The short-range lesion-dependency structure is shown for 1333 binary ischaemic (top) and 159 binary metabolic (bottom) lesion maps. Given any lesioned voxel location, the conditional probability of each of six neighbouring voxels also being affected was summed into a single vector pointing towards the direction of greatest local dependence (and therefore potential inferential distortion). The resultant voxel-wise conditional dependency vectors are 3-dimensionally rendered as arrow glyphs against orthogonal slices through a canonical white matter surface in MNI space. Larger magnitude is represented with warmer colours and larger glyphs. Ischaemic lesions show a striking pattern: lesioned voxels are strongly and systematically influenced by damage to other voxels within the proximal arterial distribution. In contrast, voxels within metabolic lesions show minimal and relatively unstructured local dependencies from which more robust lesion-deficit inferences can be drawn.

Figure 2

Covariance structure of metabolic lesions: global dependency. The long-range correlation structure is shown for 1333 binary ischaemic and 159 binary metabolic lesion maps. To avoid missing potential differential hemispheric biases, the left and right hemispheric voxels are presented separately. The correlation between every pair of grey matter voxels was binned according to degree of displacement in the x, y and z planes. Isocontours of the median correlation coefficient (rho) are presented as a function of displacement. Metabolic lesions are spatially isotropic, whilst ischaemic lesions show asymmetric elongated spatial correlations in the y and z planes especially. It is the anisotropy of the latter that distorts mass-univariate lesions-deficit analysis.

Figure 3

Metabolic lesion-deficit mapping of the components of the Wechsler Adult Intelligence Scale (WAIS): verbal IQ. Voxel-wise statistical parametric lesion-deficit maps of WAIS verbal IQ and subcomponents are three dimensionally rendered onto a canonical white matter surface in MNI space. Only voxels surviving the P < 0.05 two-tailed FWE correction for multiple comparisons are shown. Voxels are coloured according to their corresponding t-statistic, with positive associations (where hypometabolism corresponds to an impairment of cognitive scores) displayed on a red-yellow scale and negative associations displayed on a blue-green scale. Three different rotations of each map are shown per row of images next to the test labels. dACC = dorsal anterior cingulate cortex; vACC = ventral anterior cingulate cortex.

Figure 4

Metabolic lesion-deficit mapping of the subcomponents of the Wechsler Adult Intelligence Scale (WAIS): performance IQ. Voxel-wise statistical parametric lesion-deficit maps of the WAIS performance IQ and matrix reasoning subcomponent are shown. Image characteristics and abbreviations are as in Fig. 3.

Figure 5

Metabolic lesion-deficit mapping of memory. Voxel-wise statistical parametric lesion-deficit maps of individual memory tests. Image characteristics and abbreviations are as in Fig. 3.

Figure 6

Metabolic lesion-deficit mapping of fluency and affect. Voxel-wise statistical parametric lesion-deficit maps of fluency and depression (HADS). Higher depression scores are pathological, but the score has been reversed in the above images to match other psychological scores used: a positive correlation implies that hypometabolism corresponds to an impairment of affect (greater depression) and is shown on a red-yellow scale. Image characteristics and abbreviations are as in Fig. 3.

Figure 7

Metabolic lesion-deficit clinical prediction of WAIS-verbal IQ and depression. Penalized Bayesian multiple regression was used to predict WAIS verbal IQ (VIQ) and depression (HADS) reasonably well from ¹⁸F-FDG PET data. Although the model is evaluated in terms of predictive accuracy, it is interesting to compare the support for these multivariate predictions—the weighting of each voxel in the model—with the weightings assigned in the univariate case. Multivariate weightings (right) presented as the t-statistic are broadly similar to the univariate weightings (left, reproduced from Figs 3 and 5), providing further support that the univariate maps represent genuine, undistorted structure-function maps.