Cortical lesions causing loss of consciousness are anticorrelated with the dorsal brainstem

Abstract Brain lesions can provide unique insight into the neuroanatomical substrate of human consciousness. For example, brainstem lesions causing coma map to a specific region of the tegmentum. Whether specific lesion locations outside the brainstem are associated with loss of consciousness (LOC) remains unclear. Here, we investigate the topography of cortical lesions causing prolonged LOC (N = 16), transient LOC (N = 91), or no LOC (N = 64). Using standard voxel lesion symptom mapping, no focus of brain damage was associated with LOC. Next, we computed the network of brain regions functionally connected to each lesion location using a large normative connectome dataset (N = 1,000). This technique, termed lesion network mapping, can test whether lesions causing LOC map to a connected brain circuit rather than one brain region. Connectivity between cortical lesion locations and an a priori coma‐specific region of brainstem tegmentum was an independent predictor of LOC (B = 1.2, p = .004). Connectivity to the dorsal brainstem was the only predictor of LOC in a whole‐brain voxel‐wise analysis. This relationship was driven by anticorrelation (negative correlation) between lesion locations and the dorsal brainstem. The map of regions anticorrelated to the dorsal brainstem thus defines a distributed brain circuit that, when damaged, is most likely to cause LOC. This circuit showed a slight posterior predominance and had peaks in the bilateral claustrum. Our results suggest that cortical lesions causing LOC map to a connected brain circuit, linking cortical lesions that disrupt consciousness to brainstem sites that maintain arousal.


32
Defined as total number of (2mm isotropic) voxels in the binary mask in MNI152 space.

35
These were defined as lesions having at least 20 lesioned voxels on the respective side 36 of the mid-sagittal plane in MNI space. If lesions had more than 0 but fewer than 30 37 voxels on a given side, they were visually inspected to confirm hemispheric involvement.

40
We sought to capture all lesions with parenchymal supratentorial involvement in each 41 hemisphere. We defined these lesions as having >10 voxels lesioned in each identified the subset of all lesions with even one voxel involving the brainstem or 0). Lesions had to involve at least 10 cerebellar voxels or at least 5 brainstem voxels to 50 be classified as having cerebellar or brainstem involvement respectively.

53
Due to the small number of thalamic lesions, we first identified lesions with at least one 54 voxel overlapping the thalamus (Harvard-Oxford subcortical atlas mask, threshold 0).

55
We then excluded several lesions that were primarily cortical but had deep extension 56 and fewer than 5 voxels of peripheral thalamic involvement. All our lesions classified as 57 having thalamic involvement had more than 10 voxels lesioned within this thalamic mask 58 and qualitatively showed unequivocal thalamic involvement. There were two lesions with 59 more than 10 lesioned voxels but qualitatively peripheral thalamic involvement.

60
Excluding these two lesions had no effect on analyses requiring identification of thalamic 61 lesions.

76
Lesion characteristics 77 Each lesion characteristic (lesion volume, brainstem involvement, thalamic involvement, 78 posterior fossa involvement, left/right hemisphere involved, bilaterality) was tested for 79 association with LOC in an ordinal logistic regression using a proportional odds model.

80
For no predictor variable was the proportional odds assumption violated (Likelihood 81 Ratio Test of non-proportional odds, P > .2). The model including volume and 82 hemisphere had a better fit than a model containing volume or hemisphere alone (LR 83 9.3, P = .002) but was not significantly affected by the addition of bilaterality (LR 2.1, P = 84 .1).

87
We performed voxel-wise regressions between the presence of a lesion at each 88 individual parenchymal voxel and LOC, defined ordinally. We restricted the analysis to 89 voxels lesioned at least five patients ( Figure S5). VLSM studies require such thresholds 90 to avoid misplacement biases (Karnath, Sperber et al. 2017). Because there is no clear 91 theoretical reason require a five-patient minimum, we repeated the analysis with a ten-92 patient threshold. In both cases, controlling the FDR (q = .05) using Benjamini-Hochberg 93 procedure(Benjamini 1995), we found no significant voxels. We also found no significant 94 voxels when treating LOC as a binary variable and using voxel-wise Chi-squared or 95 Fisher exact tests.

97
Lesions' connectivity with coma-specific area (ROI to ROI analyses)

102
The addition of lesion connectivity to the coma-specific area improved a model including 103 lesion volume (LR 6.7, P < .01), or lesion volume and hemisphere (LR 11.2, P < .001).

104
To confirm our results were not biased by deep lesions, we excluded the 20 lesions with 105 involvement of the thalamus, brainstem, or cerebellum and repeated the regressions,

120
To find the whole-brain peaks, we performed logistic regressions at every voxel within a 121 whole brain T1 2mm MNI brain mask, again including the above covariates in each 122 regression. We used a raw P value threshold of 0.001 to identify the peaks ( Figure S1). Voxels with FWE corrected P < 0.05 in all voxel-wise analyses using PALM (no 128 covariates, lesion volume, or lesion volume and hemisphere) were used to create a seed 129 region. We computed the Fisher Z-transformed Pearson R between this seed region and 130 every other voxel within the MNI brain mask. The anticorrelated regions represent the 131 areas preferentially intersected by LOC-causing lesions ( Figure S4). We used FSL's 132 cluster function to define the coordinates of the peaks (cluster size > 50 voxels, T value 133 > 7) (Table S4).

135
Given the proximity of the seed to the 4 th ventricle, we sought to ensure our anti-136 correlated network map was not driven by CSF signal. We created a spherical ROI

143
Brainstem temporal signal-to-noise ratio (tSNR) 144 Given that our connectivity findings mapped to the brainstem tegmentum, we further 145 investigated the temporal signal to noise ratio (tSNR) within this region of brainstem to 146 ensure the presence of reliable BOLD signal in this region.

148
We calculated the voxel-wise mean tSNR from the Brain Genomics Superstruct Project       Figure S1. Whole brain peak-search Voxel-wise correlations between lesions' connectivity and LOC, in multivariate regressions including lesion volume and left-hemisphere, within a T1 2mm brain mask. At a raw P threshold of 0.001, the whole-brain grey matter peak is shown. T value -7 -12 Figure S4. Anticorrelated network Coronal slices of the network of brain regions anticorrelated (T ≤ -7) to the peak brainstem cluster ( Figure 5B). Anti-correlated network (T value) -7

Voxels lesioned in ≥ 5 patients
Anti-correlated network (4 th ventricular signal partialed out) -T value -7 -12 -12 Figure S6. Anti-correlated network, 4 th ventricular signal removed Coronal slices showing the resulting network anticorrelated (T ≤ -7) to the brainstem peak ( Figure 5B) after controlling for connectivity to the 4 th ventricle (CSF signal). The left cingulate, M1, and lateral occipital regions are diminished in significance, but the network is otherwise similar.