An Electroencephalography Connectomic Profile of Posttraumatic Stress Disorder

In the method validation study, 36 medication-free healthy civilian participants were recruited from the local Stanford community. These participants had never met lifetime criteria for DSM-defined psychiatric disorders, as assessed with the Structured Clinical Interview for DSM-IV (SCID) (22). They likewise had no history of neurological injury or head trauma and had very low levels of symptoms on self-report clinical scales (Table 1).

Participants in the clinical study (Table 2) were 201 combat-exposed U.S. veterans who took part in Operations Enduring Freedom (Afghanistan) or Iraqi Freedom/New Dawn (Iraq). Recruitment was conducted through flyers, social media, and direct contact through VA records. Military service was confirmed using government-issued military or veteran identification and Department of Defense form 214 (discharge document). A diagnosis of full or subthreshold PTSD (three of four clusters met) was made using the Clinician-Administered PTSD Scale (CAPS) for DSM-5 (23) by a licensed doctoral-level psychologist. We included full and subthreshold PTSD both because a high degree of impairment has been found in subthreshold PTSD (24, 25) and in order to include a broader range of clinical participants and avoid diagnostic discrepancies related to the changes in diagnostic criteria for PTSD across DSM-IV and DSM-5 (26–28). Diagnosis of other comorbid conditions, such as major depressive disorder, was done using the SCID. All diagnoses were confirmed in consensus clinician meetings. A history of traumatic brain injury (TBI) was determined on the basis of whether loss of consciousness occurred after a combat-related head trauma. In addition to the CAPS and the SCID, participants completed the Beck Depression Inventory (BDI) to assess depressive symptoms. All medication dosages for the PTSD sample were stable for 3 months before study enrollment, and the control sample was free of psychiatric medication.

A comprehensive computer-administered neurocognitive battery was used to assess multiple domains, emphasizing executive functions, including working memory (digit span task), inhibition (go/no-go task), and shifting (Trail Making Test, parts A and B) (29, 30). Impairments of these domains in PTSD are common (6), and we chose representative measures in each to examine the behavioral correlates of differences in functional connectivity. For additional details, see the Supplemental Methods section in the online supplement.

EEG recordings were acquired with a BrainAmp DC amplifier (sampling rate, 5 kHz; measurement range, ±16.384 mV; cutoff frequencies of the analog high-pass and low-pass filters, 0 and 1 kHz) and the EasyCap EEG recording cap with 64 sintered Ag-AgCl electrodes (Brain Products GmbH, Germany). The electrode montage followed an equidistant arrangement extending from below the cheekbone back to below the inion. Electrode impedances were kept below 5 kΩ. An electrode attached to the tip of the nose was used as the reference. Participants were seated on a comfortable reclining chair and were instructed to remain awake and let their mind wander in the eyes-closed paradigm and then to fixate on a given point in the eyes-open paradigm, each for 3 minutes. Recordings were immediately assessed for quality using a custom MATLAB (release R2014b; MathWorks, Inc., Natick, Mass.) script and were rerun if necessary. Details of our preprocessing and source localization approach are described in the online supplement. The resulting EEG data were then filtered into four canonical frequency ranges: theta (4–7 Hz), alpha (8–12 Hz), beta (13–30 Hz), and gamma (31–50 Hz).

The analytical signal, which is the complex-valued time series containing information about phase and amplitude, of each vertex of source-space was separately and iteratively orthogonalized with respect to all other vertices. Power envelopes were then calculated from each of the orthogonalized analytical time series, and the natural logarithm of these envelopes was taken to render them more normal (18). A Pearson’s correlation coefficient between the log-transformed power envelopes was obtained for each vertex pair. For comparison, connectivity matrices for the raw power envelopes, calculated from non-orthogonalized analytical time series, were similarly computed. This workflow is illustrated in Figures S1–S3 in the online supplement. Additionally, a well-annotated MATLAB script used to perform these connectivity analyses is included in the online supplement.

Connectivity was calculated among 31 regions of interest in Montreal Neurological Institute space (see Figure S4 in the online supplement), derived from an independent parcellation of resting-state fMRI connectivity using independent components analysis from 38 healthy subjects applied in a previous study (31). This was done to ground our EEG regions of interest in a commonly used, fMRI-derived definition of the major cortical connectivity networks while maintaining region-of-interest sizes large enough to be transformed for use in the lower-resolution electrophysiological context (10, 32).

For each region-of-interest pair, the Fisher z transformations of the Pearson correlations of each pair of vertices within each of the regions of interest were averaged to determine the connectivity of that region-of-interest pair. As a result, 465 unique region-of-interest pairs (excluding self-connectivity) were computed in each of the four frequency bands, in each of the two resting paradigms (eyes open and closed), and in each of the two computational methods (plain and orthogonalized).

Mean region-of-interest pair connectivity values were analyzed in parallel linear mixed-effects (LME) models with PTSD diagnosis as a categorical predictor, collection site as a covariate, and random intercept for participant. The p values for this analysis were corrected using a single false discovery rate (FDR) adjustment (p_{FDR, 1860 comparisons}<0.05; Benjamini-Hochberg method) that included all regions of interest and frequency bands. Outliers in connectivity measures exceeding the median ±3 times the interquartile range were excluded (mean=2.7 subjects, SD=2.6).

To determine whether subjective clinical symptoms or behavioral measures of cognitive domains mapped onto measures of connectivity, we used the surviving region-of-interest connections from the above LME model as response in another LME model including site as covariate and random intercept for participant. The cognitive domain measures, CAPS total score, CAPS subscale scores, and BDI total score were set as separate predictors. Nonresponses or timed-out responses in the behavioral data were excluded (four, three, and two subjects in the working memory, inhibition, and shifting tasks, respectively).

Figure 2 shows network connectivity patterns attained in our validation sample using power envelope connectivity. As predicted from previous work developing and validating this method (18, 19), the expected connectivity patterns emerged only after orthogonalizing power envelopes to suppress zero-phase-lag volume conduction. By seeding various constituent regions of interest of the major cortical networks and calculating the t test for correlation ≠ mean brain-wide correlation (p_{FDR, 3000 vertices}<0.05, two-tailed), connectivity patterns generally consistent with those observed in fMRI studies emerged—for example, the visual, dorsal attention, frontoparietal control, somatosensory, default mode, and ventral attention networks (Figure 2). Greater intranetwork connectivity as compared with extranetwork connectivity (Table 3) is seen in all networks in various frequency bands (t test for mean of region of interest to within-network vertices ≠ mean of region of interest to extranetwork vertices, p_{FDR, 24 network-frequency band pairs}<0.05, two-tailed). As expected, early sensory networks (visual and somatosensory) are highly intraconnected across nearly all frequency bands, and the default mode network demonstrates canonical intraconnectivity in the alpha and beta bands. Figure S5 in the online supplement illustrates more canonical distributed topographies achieved by calculating the t statistic of mean region-of-interest connectivity >0 rather than a brain-wide mean. Similarly, Table S1 in the online supplement describes the intranetwork minus extranetwork connectivity obtained by this method, which produces the same significant network-frequency band pairs as those listed in Table 3. Diffuse extranetwork connectivity is observed in most network-frequency band pairs, likely attributable in part to network overlap unresolved by EEG’s lower spatial resolution (e.g., the apparent overlap of the inferior parietal lobule and angular gyrus in the default mode and frontoparietal control networks).

Next, we tested, at each of the four carrier frequencies, whether power envelope connectivity was perturbed in participants with PTSD. Seventy-four region-of-interest pairs survived correction for multiple comparisons (Figure 3A). These were found exclusively in the eyes-open condition and predominantly in the theta carrier frequency. Brain region pairs demonstrated positive connectivity in the control group and underconnectivity in the PTSD group; this contrast is illustrated among selected regions of interest in Figure S6 in the online supplement. In all cases, these represented hypoconnectivity in the PTSD group. The majority of significant region-of-interest pairs were connected to the orbital and anterior middle frontal gyri. The whole-brain FDR-corrected vertex-wise PTSD hypoconnectivity maps for these two regions are shown in Figure 3B and 3C. As can be seen, hypoconnectivity is particularly pronounced across the lateral and medial frontal cortex as well as the inferior parietal lobe.

In contrast to these results, similar group comparisons of regional power spectral density in each of the frequency bands, using the same false discovery rate correction, failed to yield any significant findings (data not shown). Hence, our results are specific to power envelope connectivity, almost exclusively within the theta frequency band, and are not evident if one examines only the mean spectral power of that frequency band (theta included). Moreover, consistent with expectations about the importance of orthogonalization during the calculation of power envelope connectivity, an analysis similar to that described above conducted on non-orthogonalized power envelopes failed to yield any significant group difference results (data not shown). Global mean connectivity was also examined in both clinical groups, and no significant differences were found in any band or paradigm.

To assess the detection power of the power envelope connectivity as a potential biomarker, we used the orthogonalized connectivity values as features to train a machine-learning classifier for distinguishing PTSD patients from trauma-exposed control subjects. Classification performance was evaluated using a leave-one-out cross-validation procedure. The receiver operating characteristic curve (ROC) (see Figure S7 in the online supplement) shows promising classification performance (area under the curve [AUC]=0.898, sensitivity=80.2%, specificity=84.9%). To further confirm the biomarker’s generalization capability, we also performed a 10×10-fold cross-validation procedure. The ROC analysis shows that our discovered biomarker still provided good classification performance (AUC=0.813, sensitivity=76.3%, specificity=74.9%; see the Supplemental Methods section in the online supplement), demonstrating its generalization capability for identifying PTSD patients.

To understand the relevance of the heavily prefrontal and chiefly theta-based connectivity abnormalities in cognition in PTSD, we correlated performance on tasks of executive functions (working memory, inhibition, and shifting) with connectivity in the significant region-of-interest pairs that emerged from the analysis described above. The PTSD group performed significantly worse than the control group on the digit span task (Figure 4A). Across both groups, individual differences in digit span performance furthermore correlated with individual differences in power envelope connectivity for 25 region-of-interest pairs of the 74 that showed a significant group difference, including within-network pairs in the dorsal attention, frontoparietal control, and ventral attention networks (Figure 4B). No significant group differences were found for other cognitive measures.

We additionally tested whether the observed connectivity abnormalities were related to clinical features of PTSD. However, measures of symptom severity on the CAPS total and subscale scores, as well as on the BDI, did not significantly correlate with connectivity among the region-of-interest pairs identified in the group contrast above. Occurrence of comorbid major depression or TBI also did not account for significant differences in connectivity among the region-of-interest pairs identified above. Similarly, while 26.4% of the PTSD group used daily or as-needed psychotropic medications (see Table 2), use of psychotropic medication did not account for significant differences in connectivity.