Association of Generalized Anxiety Disorder With Autonomic Hypersensitivity and Blunted Ventromedial Prefrontal Cortex Activity During Peripheral Adrenergic Stimulation

Key Points Question Do individuals with generalized anxiety disorder (GAD) show abnormal physiological, perceptual, or neural responses during peripheral β-adrenergic stimulation that may indicate interoceptive dysfunction? Findings In this crossover randomized clinical trial, female patients with GAD exhibited hypersensitivity to adrenergic stimulation as well as greater interoceptive sensation and diminished ventromedial prefrontal cortex activity compared with healthy participants. Meaning This study provides evidence of dysfunctional autonomic and central nervous system contributions to the pathophysiology of GAD and suggests that the ventromedial prefrontal cortex may be a treatment target.


eMethods 1. Participants
One hundred thirty-six participants were assessed for eligibility, following study recruitment from outpatients receiving services at the Laureate Psychiatric Clinic and Hospital as well as via newspaper, radio, flyer, and social media advertisements in the Tulsa metropolitan area. Those with current/prior schizophrenia, bipolar disorder, or obsessive-compulsive disorder were excluded. Participants were allowed to continue use of selected psychotropic medications but were instructed not to take any analgesic or other over the counter (OTC) medication within 48 hours of the experimental session and no benzodiazepines, beta-blockers, or tetrahydrocannabinol (THC) containing products within 24 hours of scanning. No participants tested positively for THC or benzodiazepines on the day of scanning. They were further instructed not to have nicotine on the day of the experimental session and to limit their caffeine consumption (i.e., avoid caffeine it was not a daily routine for the participant, and if it was, to reduce their usual amount by half).
Eleven candidates declined to participate and the remaining 125 provided informed written consent. Seventeen participants withdrew prior to the intervention for various reasons (scheduling difficulty, incomplete baseline assessment, or study refusal). Thirty-five additional participants were excluded due to contraindicated medical conditions. Medical contraindications included: 1) (a) cardiac disease (e.g. bradycardia <40 bpm, any history of diagnoses EKG abnormalities or evidence of 12-lead EKG abnormalities during baseline visit or evidence of frequent premature ventricular contractions during baseline visit), (b) respiratory disease (asthma, pneumonia), (c) endocrine disease (hypothyroidism, diabetes), (d) neurological disease (seizure disorder or reporting a seizure within the last year, head injury with sustained loss of consciousness and memory loss), (e) pain disorder (f) history of hepatic or renal failure, (g) gastrointestinal, hematologic, rheumatologic, or metabolic disease, 4) pregnancy as verified by urine test, 5) intrauterine device (IUD) posing a scanning risk, 6) ages 18 to 55 years, 7) excessive BMI (BMI > 35) preventing scanner entry, 8) MRI contraindications including: (a) cardiac pacemaker, (b) metal fragments in eyes/skin/body (shrapnel), (c) aortic/aneurysm clips, (d) prosthesis, (e) by-pass surgery/coronary artery clips, (f) heart valve replacement, (g) shunt (ventricular or spinal), (h) implanted electrodes, (i) MRI contraindicated metal plates/pins/screws/ wires, or neuro/bio-stimulators (TENS unit), (j) persons with a history of professional metal working/welding who showed evidence of intraocular metal on orbital xray, history of eye surgery/eyes washed out because of metal, (k) inability to lie still on one's back for a 1-2 hours, (l) prior neurosurgery, (m) tattoos or cosmetic makeup with metal dyes, (n) unwillingness or inability to remove body piercings, 9) medications that affect the hemodynamic response (e.g. acetazolamide, excessive caffeine intake > 1000 mg per day, 10), noncorrectable vision or hearing problems, or 11) vital sign abnormalities during the screening visit (e.g. systolic blood pressure >160 mmHg, diastolic blood pressure >100 mmHg, resting pulse >100 beats per minute).
Of the 73 allocated to the intervention, 72 completed it. One GAD participant withdrew during the infusions due to panic anxiety. Five participants were excluded due to excessive head motion resulting in low data quality (four HCs and one GAD), eight HCs were excluded during the matching process based on age and BMI, and one GAD was excluded for having a BMI too high to appropriately match with a healthy comparator. All participants were financially compensated for their participation.

eMethods 2. Experimental Protocol
Isoproterenol was obtained from Valeant Pharmaceuticals, Laval, QC, Canada and infusions were prepared into 3 mL boluses by a pharmacist who was unblinded. The isoproterenol infusion administration procedure mirrored our previous fMRI cross-over protocol in healthy volunteers 1 , with infusions administered by a nurse seated beside the scanning bed inside of the MRI room who was blinded to infusion condition. For safety monitoring purposes, cardiac rhythm was also recorded continuously using two MRI-compatible ECG leads (lead I and II configuration) (GE Healthcare, Waukesha, WI, USA). These rhythms and vital signs were monitored continuously by the nurse delivering the infusions. As an additional safety precaution, condition order was available to the research assistant sitting in the control room, who was blinded to the study hypotheses.
The randomized sequences for dose order were individually predetermined via a random number generator prior to beginning study sample recruitment. Upon recruitment of each participant, a 3 rd party registered nurse (i.e., uninvolved in infusion administration or data collection) selected a randomization code from a set of randomizations generated by the senior study author. The 3 rd party nurse then arranged the bolus doses in the predetermined order, and after obtaining verification of this order by a second 3 rd party nurse (or nursing assistant), covered the labels. Selection of a randomization order for a given participant was not determined by group membership and occurred in sequential order from the list of randomization orders.
On arrival participants ate a 300 Calorie snack. Each participant was led through a training session approximately one hour before fMRI scanning. During this session, the participant was instructed that they would receive both isoproterenol and saline infusions at different points during the scan. Participants were informed that "you may notice an increase in your heartbeat sensations, and/or may notice increase in your breathing sensations" during the isoproterenol infusions. Also, to familiarize participants with the experience of the infusion setup and isoproterenol-induced sensations, prior to the scan, participants received two practice bolus infusions (saline and 1 microgram, g) delivered by the nurse in a separate room near the MRI scanning suite. These infusions followed the same time course as those administered during the scanning session and required participants to provide sensation ratings using a dial. We administered a 1 g dose during the practice infusion to avoid a familiarity effect during the subsequent infusion scans, and to ensure that participants received a large enough dose that they were likely to perceive, based on our prior studies [2][3][4] . Following the scan session participants ate a 1000 Calorie meal.

eMethods 3. Primary and Exploratory Behavioral and Physiological Data Analysis
All statistical tests were performed in R (version 3.5.1; R Core Team, 2016). Linear mixed-effects (LME) regression models were performed using the lme4 package 5 , and statistical degrees of freedom were estimated by Satterthwaite's approximation via the lmerTest package 6 . We assigned the reference levels for the group, dose, and epoch variables to be the healthy comparator group, the saline condition, and the baseline epoch, respectively. We estimated the chronotropic dose 25 (CD25) by first performing a withinsubject linear regression, by ISO dose, on the maximum average heart rate occurring over any rolling 5 second window during the peak and early recovery periods. We used windowed means as opposed to point estimates to reduce potential measurement error due to outliers, as in our previous studies 7 . To provide a baseline for calculating the CD25, we also obtained rolling 5 second windows for the maximum heartrate occurring during the corresponding 40 second period of a resting state scan run before the task. We then calculated each participant's CD25 estimate using the formula provided in Mills et al. 8 , by adding the 25-beat target increase to the resting state value, subtracting the intercept of the linear regression, and dividing the subsequent value by the regression slope. Interoceptive detection was assessed via a chi-square test for independence in R software which counted, for each dose, how many GADs and HCs increased the dial above a threshold of fifteen out of one hundred (this dose detection threshold was chosen to account for inadvertent responses due to the MRI compatible device's sensitivity profile). Differences in cardiorespiratory interoceptive accuracy were assessed by calculating cross-correlations between continuous heart rate and dial ratings for each individual using custom MATLAB script (as in 1 ) prior to LME model generation. All physiological and subjective dial responses were down sampled to 1 Hz prior to calculating cross-correlations. We calculated cross correlations for each participant from mean centered dial ratings separately with instantaneous changes in HR and respiratory volume variability (RVV) occurring over the 120 s interval following infusion onset. For each dose, we calculated the cross-correlation with no lag time, the maximum cross-correlation allowing for temporal shifting, and the lag time associated with the maximum cross-correlation. Group and dose differences for each of these coefficients (HR and RVV) were assessed by LME models in an identical manner to the retrospective ratings. Exploratory Pearson correlations for vmPFC cluster PSC and ROI FPC scores with real-time and retrospective cardiorespiratory or anxiety intensity were conducted using the correlation package in R 9 and corrected for multiple comparisons via the Hommel method 10 .

eMethods 4. MRI Data Acquisition
MRI scans were acquired in a GE MR750 3T scanner with an 8-channel head coil. Simultaneous EEG recordings were also collected but were not included in the current analysis. We acquired anatomical images via a T1-weighted magnetization-prepared rapid gradient echo (MPRAGE) sequence with sensitivity encoding (SENSE 11 ) that lasted 5 min and 40 s. MPRAGE parameters were: 190 axial slices, slice thickness = 0.9 mm, TR/TE = 5/2.012 ms, FOV = 240 x 192 mm 2 , matrix size 256 x 256, flip angle = 8⁰, inversion time = 725 ms, SENSE acceleration R = 2, with a sampling bandwidth of 31.2. A 240 s, single-shot gradient echo planar imaging (EPI) sequence covering the whole-brain was used for each fMRI scan. In this EPI sequence we obtained axial 39 slices, 2.9 mm thick, with no gap and a voxel size of 1.875 x 1.875 x 2.9 mm 3 . Additional parameters were TR/TE = 2000/27 ms, FOV = 240 x 240 mm 2 , flip angle = 78⁰, SENSE acceleration R = 2 with a 96 x 96 matrix.

eMethods 5. fMRI Preprocessing
The first 4 EPI volumes were dropped to allow field stabilization. BOLD signal was scaled to percent change from the time-course mean for each voxel. With respect to motion artifacts, poor quality volumes were censored using interpolation if they exceeded 0.3 mm of mean motion or if >20% of voxels were found to be outliers using AFNI's 3dtoutcount. Runs were discarded if >20% of volumes were censored. One run was discarded for each of three HC and two GAD participants. We also checked for potential differences in head motion for each of the infusion conditions and found no group differences.

eMethods 6. Whole-brain fMRI Group Analysis
We conducted whole-brain analysis using AFNI's 3dttest++ program and applied the ETAC (Equitable Thresholding and Clustering) option to estimate significant cluster sizes corresponding to a 5% false positive rate at a p < 0.001 voxel-wise threshold 12 . We chose this method to reduce arbitrary judgment in selection of the uncorrected p-value threshold 13 .

eMethods 7. Longitudinal time series analysis with regions of interest
Multivariate sparse functional principal component analysis (mSFPCA) was performed to examine how linear and non-linear temporal dynamics of isoproterenol's effects on BOLD response, heart rate, and sensation rating trajectories. This technique characterizes the major modes of smoothed variation around the smoothed mean trajectories as FPC coefficients at the subject level using Hamiltonian Markov chain Monte Carlo (MCMC) resampling of the posterior distribution of model parameters in a Bayesian framework 14 . To assist in functional interpretation of our whole-brain results, we generated separate post hoc mSFPCA models for extracted mean PSC from spherical ROIs (5 mm radius) derived from conceptually relevant fMRI meta-analyses reporting clusters in coordinate space near to ours. Seeds for the relevant regions of interest (ROIs) within the vmPFC were obtained from meta-analyses of fMRI studies related to valence processing 15 (x=-3, y=39, z=0), sympathetic autonomic control 16 (x=-4, y=38, z=-12 and x=8, y=42, z=-6), and selfprocessing 17 (x=2, y=53, z=7). FPC coefficients for HR, dial, and BOLD ROI trajectories were then entered into LME models with group and dose factors to assess how these factors affect those trajectories. The mSFPCA approach we employed uses sets of eigenfunctions that are allowed to correlate across dependent variables but remain independent with regard to each observation within a variable 18 . Using spline basis functions, each individual's response was modeled in a multivariate fashion via a matrix of spline coefficients characterizing the major modes of variation as a set of principal components functions (we chose two) around the smoothed mean trajectory. mSFPCA models were created in R using the MASS 19 , Matrix 20 , pracma 21 , and splines 22 packages. mSFPCA was conducted on HR and dial responses after converting the timescale of these variables to 2 second intervals, with the first four intervals dropped to match the BOLD timeseries given the TR of 2 second and the initial 4 volumes being dropped to allow the magnet to reach a steady-state. We generated mean and PC functions by implementing two internal knots yielding model fits via cubic B-spline basis functions. We ran 1,100 MCMC simulations with 100 of those discarded as burn-in. eFigure 1. Infusion epochs used in the isoproterenol fMRI analysis.
For illustrative purposes, infusion epochs are superimposed on the average time course of heart rate responses during the saline, 0.5 and 2 µg isoproterenol infusions in the healthy comparator sample only. The data shown in the figure includes the complete 43 healthy individual sample tested prior to exclusions for head motion (n=5) and age and BMI group matching (n=9). bpm = beats per minute, µg = micrograms, Sal = saline.

eFigure 2. Subjective Response Assessment
A) Throughout each scan, participants reported their perceived cardiorespiratory sensation intensity ranging using a visual analog scale with a baseline of 0 and maximum of 10 (recorded units of measurement ranged from 0 to 100). B) Immediately after each infusion scan participants provided retrospective ratings of their heartbeat sensation intensity, breathing sensation intensity, anxiety, and arousal felt during the infusions, using an 11-point scale ranging from 0 to 10.

B)
A) eFigure 5. Group difference in left parietal cortex activity.
The GAD group showed significant attenuation relative to the HC group at a p<0.001 voxel-wise threshold during the peak response period for only the 0.5 μg dose. eFigure 6. Maximum cross-correlations between heart rate and dial response by dose in unmedicated GAD relative to HC.
When medicated GAD participants were removed the mean correlation between HR and perceived cardiorespiratory intensity was significantly higher at the 0.5 μg dose in the unmedicated GAD group relative to HCs at the level of the maximum cross-correlation. Error bars = SE. eFigure 9. VmPFC PSC correlations with cardiorespiratory and subjective variables during the peak effect at 0.5µg isoproterenol Scatterplots with correlation slope line and standard deviation band for relationships between percent signal change extracted from the whole-brain vmPFC selected cardiorespiratory and self-report variables. After Bonferroni-Holm correction for multiple comparisons, significant inverse relationships were seen between PSC and A) change in heart rate from pre-infusion baseline as well as post-infusion self-report of both B) beartbeat intensity and C) respiratory intensity. Negative PSC relationships not surviving Bonferroni-Holm correction were seen for D) perceived cardiorespiratory intensity, though this remained a trend, and E) self-reported anxiety. The relationship with F) CD25 was not significant. CD25 = chronotropic dose 25 eResults 1. Respiratory Volume Variability (RVV) We observed dose-related increases in RVV during the peak period but no interaction with group (eFigure 3 and eTable 5), suggesting that group differences in perceived interoceptive intensity were driven preferentially by cardiac signals.

eResults 2. Interoceptive Awareness: Detection and Accuracy Measures
A chi-square test to assess group differences in detection rate of cardiorespiratory changes via increased dial ratings was not significant (χ(1)=2.54, p=0.111; (eFigure 4). Nor were there clear differences in interoceptive accuracy, as assessed via the crosscorrelation of HR with dial responses. The maximum cross-correlation was the only metric showing a trend toward increased accuracy for GADs during a combined peak + early recovery period at 0.5μg (b=0.09, SE=0.05, t(281.53)=1.79, p=0.075; eTables 6 and 7).

eResults 3. Reanalysis of Primary Outcomes with Medicated Participants Removed
A reanalysis with the six medicated patients removed revealed a significant difference in age between groups (t(38.28), p<0.047) with GADs being older (M=27.96 years) than HCs (M=24.38 years); age was included as a covariate in the subsequent analyses (as it was in the original models). We observed an additional significant result of higher HR for those with GAD compared to HCs at the 2mcg dose during the peak period (β=3. 76 . Results relating to BMI, CD25, and all retrospective ratings were unchanged with respect to statistical significance and nearly identical to the previous analysis.

eResults 4. Longitudinal Time-Series Analysis
The mSFPCA LME revealed an effect of group at 0.5μg for the self-processing ROI (β=0. 23

eDiscussion 1. Study Limitations
Our focus on females with GAD was based on the fact that females outnumber males on a 2:1 basis 23 . Future research will need to establish whether the current results may extend to males. Recent studies have suggested a divergence of symptom clusters in GAD females vs. males such that females are more likely to report somatic symptoms whereas males are more likely to report mood and motor symptoms 24 . Thus, it is possible that the findings observed in the current study may represent a subtype expression of the disorder specific to females. Although we allowed participants who were taking psychotropic medications to enter the study, we do not think this impacted our findings, for several reasons: 1) they were stably medicated, 2) we would have expected to see disrupted action across all doses and epochs, and 3) the main effects of hypersensitivity and blunted vmPFC signal remained even after removing the medicated participants in follow-up analyses. Finally, given the high degree of comorbidity in GAD and our diagnostic focus, future work will need to test this paradigm in dimensionally and categorically defined samples (as in 25 ) to see if a similar pattern of vmPFC hypoactivity during adrenergic stimulation is evident across other anxiety disorders.