key: cord-104055-47ren7ie
authors: Lutkenhoff, Evan S.; Nigri, Anna; Sebastiano, Davide Rossi; Sattin, Davide; Visani, Elisa; Rosazza, Cristina; D’Incerti, Ludovico; Bruzzone, Maria Grazia; Franceschetti, Silvana; Leonardi, Matilde; Ferraro, Stefania; Monti, Martin M.
title: EEG power spectra and subcortical pathology in chronic disorders of consciousness
date: 2020-09-01
journal: bioRxiv
DOI: 10.1101/695288
sha: 
doc_id: 104055
cord_uid: 47ren7ie

Objective To determine (i) the association between long-term impairment of consciousness after severe brain injury, spontaneous brain oscillations, and underlying subcortical damage, and (ii) whether such data can be used to aid patient diagnosis, a process known to be susceptible to high error rates. Methods Cross-sectional observational sample of 116 patients with a disorder of consciousness secondary to brain injury, collected prospectively at a tertiary center between 2011 and 2013. Multimodal analyses relating clinical measures of impairment, electroencephalographic measures of spontaneous brain activity, and magnetic resonance imaging data of subcortical atrophy were conducted in 2018. Results In the final analyzed sample of 61 patients, systematic associations were found between electroencephalographic power spectra and subcortical damage. Specifically, the ratio of beta-to-delta relative power was negatively associated with greater atrophy in regions of the bilateral thalamus and globus pallidus (both left > right) previously shown to be preferentially atrophied in chronic disorders of consciousness. Power spectrum total density was also negatively associated with widespread atrophy in regions of the left globus pallidus, right caudate, and in brainstem. Furthermore, we showed that the combination of demographics, encephalographic, and imaging data in an analytic framework can be employed to aid behavioral diagnosis. Conclusions These results ground, for the first time, electroencephalographic presentation detected with routine clinical techniques in the underlying brain pathology of disorders of consciousness and demonstrate how multimodal combination of clinical, electroencephalographic, and imaging data can be employed in potentially mitigating the high rates of misdiagnosis typical of this patient cohort. Search terms disorders of consciousness, subcortical pathology, EEG, MRI.

Lutkenhoff, Owen, & Monti, 2017) , there are no data directly connecting the patterns 14 of EEG power spectra and subcortical damage in long-term DOC patients, a gap which 15 is not only problematic for the clinician's interpretation of the observed EEG data, but 16 also hampers our ability to monitor, through an inexpensive, bedside, repeatable 17 technique interventions and their effects. In what follows, we address, in a large cohort 18 of patients with chronic DOC, the heretofore untested relationship between observed 19 electrocortical rhythms, patterns of subcortical brain atrophy (including thalamus, 20 brainstem, and basal ganglia), and clinical measures of awareness and arousal. (Giacino, Kalmar, & Whyte, 2004)), (ii) neurophysiological evaluation, including 10 resting EEG, and (iii) neuroradiological assessment, including MRI (see Table 1 ). 11 Insert Table 1 about here 12 The acquisition of both resting EEG and structural MRI datasets constituted the 13 inclusion criteria for the present study. Experienced raters independently assessed each 14 patient 4 times with the Italian version of the CRS-R (Sacco et al., 2011) . The best- 15 recorded performance was used to classify patients as VS or MCS. As described below, 16 55 patients were discarded due to the low quality of the MRI data, following previously 17 established procedures (Lutkenhoff et al., 2015) . Specifically, 24 datasets were excluded 18 because of excessive movement artifacts, 16 datasets were excluded due to software 19 failure in segmenting subcortical structures, 13 datasets were excluded due to poor 20 quality in the estimation of normalized brain tissue volume, and 2 subjects were 21 excluded due to large regions of signal dropout artifacts (e.g., implants) preventing tissue segmentation (see Figure S1 ). Importantly, as discussed below, exclusions do 1 not bias the analyzed sample as compared to the full cohort (see also Of these, 33% (n = 20) had a traumatic brain injury (TBI) and 67% (n = 41) had non-5 traumatic etiology (non-TBI). In particular, among non-TBI patients, 30% (n = 18) 6 suffered from anoxic brain injury and 37% (n = 23) from haemorrhagic and/or ischemic 7 brain injury. The median disease duration at the time of the study was 24 months (range = 8 5-198 months). (See Tables 1 and S1 ). The local Ethics Committee approved all aspects of 9 this research and written informed consent was obtained from the legally authorized 10 representative of the patients prior to their inclusion in the study. 11 Data acquisition and analysis 12 EEG data acquisition and processing 13 Patients underwent polygraphic recordings between 2 pm and 9 am on the following 14 order to prevent signal degradation due to the prolonged recording and to maintain, as 7 much as possible, the same reproducible characteristics between recordings in different 8 patients. The selected epochs were filtered (1-70 Hz, 12 db/octave), followed by a 50 9 Hz notch filter to suppress the noise of the electrical power line, reformatted against 10 the linked ear-lobe reference. In order to remove blink-artifacts, we applied an ICA- 11 artifact rejection algorithm. Then, the selected EEG activity was divided into 90 non- 12 overlapping 2 s segments and analyzed using the fast Fourier transform. Absolute total 13 power and relative power were evaluated in the delta (1-4 Hz), theta (4-8 Hz), alpha 14 (8) (9) (10) (11) (12) (13) , beta (13-30 Hz) and gamma (>30 Hz) bands, and averaged within each 15 EEG channel. 16 

MRI data acquisition and processing 18 Neuroimaging data were obtained with a 3T MRI machine (Achieva, Philips Healthcare 19 BV, Best, NL; 32-channel coil reconstructed into 3-dimensional vertex meshes, as depicted in Figure 1 . In addition, 7 the normalized brain volume (NBV), a measure of global brain atrophy including 8 white and gray matter volume, was calculated for each patient using FSL SIENA months post-injury as fixed variables and subjects as the random variable. As 13 described below, the significant interaction between EEG features and diagnosis was 14 followed up with one mixed-model analysis per each EEG feature (using the same 15 model, albeit without the EEG feature as repeated variable). Individual mixed-models 16 were followed up with pairwise post-hoc comparisons between diagnostic groups (i.e., 17 VS, MCS-, MCS+), with Šidák correction for multiplicity. 18 19 EEG -MRI analysis 20 In the second analysis, we related EEG spectral features to subcortical shape measures. positively on the total power for each electrode (henceforth, total power component). 10 Finally, the last three components appeared to capture diffuse statistical covariance 11 between electrodes, although with a preference for loading, respectively, positively on 12 the delta band and negatively on the alpha band in right hemispheric electrodes 13 (henceforth, 18 variables, in a general linear model predicting local shape patterns (e.g., atrophy). Sex, 19 age, time-post-injury, etiology (i.e., TBI vs non-TBI), were included as covariates, 20 along with NBV (to ensure that observed tissue displacement reflect local subcortical 21 shape changes independent of overall brain atrophy ( In this analysis, we related the patients' behavioral presentation, as captured by the 7 CRS-R subscales, with subcortical atrophy. Because of significant correlations 8 between the subscales of the CRS-R (i.e., the desired independent variables), 9 behavioral data were entered into a PCA performed analogously to the one described 10 above. The analysis returned 3 components with an eigenvalue greater than 1, 11 cumulatively explaining 69.57% of the variance. The three components loaded on, 12 respectively, the auditory, visual, and arousal subscales (henceforth, audio-visual- 13 arousal component), the motor subscale (henceforth, motor component), and the 14 oromotor and communication subscales (henceforth, oromotor-communication 15 component). As in the previous analysis, the three components were entered as 16 independent variables in a general linear model predicting subcortical shape, along with 17 the same covariates described above. Group-level significance was assessed identically 18 to the previous analysis. 19 20 Predicting DOC level from EEG spectral features 21 Finally, we employed a binary logistic regression to evaluate the degree to which 

EEG results 9 The mixed-model analysis revealed a significant interaction (F (10, 1309.055) = 16.599, 10 p < 0.001) between diagnostic group (i.e., VS, MCS-, MCS+) and EEG features (i.e., 11 total power, delta, theta, alpha, beta, gamma frequency bands; see Figure 2 and Table   12 1), along with a significant main effect of diagnosis (F (2, 4389.124) = 5.158, p = Predicting DOC level from EEG spectral features 2 As shown in Figure 5a and Table 2 , diagnosis (i.e., VS vs. MCS) was predicted 3 significantly better by the model including EEG components, overall brain atrophy 4 (including both white matter and gray matter), and demographic components (i.e., age, and specificity (sens/spec; 0.79, 0.81, respectively) than both other models (AUC = 1 0.78, sens/spec 0.58/0.78 and AUC = 0.67, sens/spec 0.29/0.89 for the 2 demographics and atrophy and the demographic only models, respectively). 3 Notably, the contribution of increasingly complex models (i.e., adding brain 4 atrophy and EEG components) is to increase the model's sensitivity to MCS (at the 5 cost of a loss of specificity). Insert Table 2 about here 8 Finally, in terms of individual variables, as shown in Figure 5b and could correctly classify patients across the conscious/unconscious line (as behaviorally 12 defined) with ∼ 87% success, leveraging on demographic information (i.e., sex, age), 13 overall brain atrophy, and EEG features (i.e., total power and θ /δ components). It is 14 also noteworthy that in our data combination of EEG and MRI data is additive in terms 15 of enhancing discriminability across groups, strengthening the idea that multimodal 16 approaches are a desirable way of assaying different -and complementary -aspects of 8 Finally, in evaluating the above results, the reader should be mindful of some 9 limitations. First, our results are skewed by survivor bias effects; we might thus be 10 representing a spectrum of impairment which, while severe, excludes the even greater 11 damage present in patients who do not survive later than a year post injury. The same is 12 true with respect to the fact that 55 datasets had to be excluded from the collected sample. 13 While it would have been ideal to be able to retain more of the sample, conventional 14 quality control limited our ability to analyze the full dataset. Nonetheless, the final 15 analyzed sample is similar to that of other MR-based work in the field (e.g., ( such inferences from a small subset of our data, it is conceivable that MCS-patients 10 could have a more "rigid" rhythmic EEG less susceptible to variations as compared to 11 MCS+ patients. Third, due to significant correlations across channels within and across 12 power bands, in order to perform the regression analyses presented above, we had to 13 first reduce the independent variables by means of a PCA. While conventional, it does 14 affect the interpretation of our results in as much as we cannot directly assess whether 15 the effects we report in mixed component (e.g., the β /δ component) are principally due 16 to either frequency or to their combination. Nonetheless, each of the three ratio 17 components correlates strongly with the "raw" ratio of each pair of frequencies 18 (calculated by taking the ratio of the average relative power across all channels; 19 specifically, r=0.97, r=0.87, and r=0.85 for β /δ, α /δ, and θ /δ, respectively) as well as 20 with each numerator and denominator variable (albeit numerically more with the 21 nominator for β /δ and α /δ components; see Table S2 ). This suggests that our results can be reasonably interpreted as reflecting actual ratios in relative power. Third, 1 gamma frequencies are known to often contain residual muscle artifacts. We decided 2 to keep them in the analysis mainly because, even if they do contain artifacts, 3 including this component still contributes to explaining variance in the signal. Had we 4 not included it, any variance across patients due to motion would have de facto been 5 subsumed by the unexplained variance term. In this sense, our approach is analogous 6 to the conventional inclusion of motion parameters in functional MRI analysis. Fourth, 7 while we report associations between brain damage in subcortical regions and EEG 8 spectral features, this does not necessarily imply that the pinpointed areas are 9 themselves the generators of specific oscillatory rhythms at rest. 10 Finally, it should also be pointed out that the present work did not include an 11 additional independent sample to confirm the classification results, thus inviting 12 caution in the extrapolation of the results towards new cohorts. As multi-site efforts 13 increase, larger cohorts of high-quality data will permit full application of such 14 approaches. 15 In conclusion, this work bridges different levels of analysis of patients surviving 16 severe brain injury, uniting DOC brain pathology ( Potential Conflicts of Interest 12 The Authors declare no conflicts of interest. 13 14 Ethical standards 15 The authors assert that all procedures contributing to this work comply with the ethical 16 standards of the relevant national and institutional committees on human experimentation 17 and with the Helsinki Declaration of 1975, as revised in 2008. 18 19 Funding sources 20 This work was supported by the Italian Ministry of Health (research grant RF-2013- 

Table S1………………………………………………………………….……………….38 Table S2 .………………………………………………………………………………….39 

Dual function of thalamic low-vigilance state oscillations: rhythm-11 regulation and plasticity

Intrinsic functional connectivity differentiates minimally conscious 15 from unresponsive patients

Human consciousness is supported by dynamic complex patterns of brain signal 19 coordination

Early detection of consciousness in patients with acute severe traumatic brain 22

Thalamic and extrathalamic mechanisms of consciousness after severe brain 11 injury

Thalamic atrophy in antero-medial and dorsal nuclei correlates with six-month 14 outcome after severe brain injury

Optimized brain extraction for pathological brains (optiBET)

The subcortical basis of outcome and cognitive impairment in 21 TBI: A longitudinal cohort study

Thalamo-frontal connectivity mediates top-down cognitive functions in disorders of 13 consciousness

Willful modulation of brain activity in disorders of consciousness

Functional connectivity of EEG is subject-specific, associated with phenotype, and 20 different from fMRI

The neural correlates of lexical processing in disorders of and appearance for subcortical brain segmentation

Basal ganglia control of sleep-wake 9 behavior and cortical activation

Multimodal study of default-mode network integrity in 13 disorders of consciousness

Significance of multiple neurophysiological measures in patients 17 with chronic disorders of consciousness

Sleep patterns associated with the severity of impairment in a large cohort of 21 patients with chronic disorders of consciousness

Accurate, robust, and automated longitudinal and cross-sectional brain change 11 analysis

Diagnostic precision of PET imaging and functional MRI in 14 disorders of consciousness: a clinical validation study

Default network connectivity reflects the level of consciousness in 18 non-communicative brain-damaged patients

Disentangling 21 disorders of consciousness: Insights from diffusion tensor imaging and machine learning