Open-Label Placebo Injection for Chronic Back Pain With Functional Neuroimaging

Key Points Question What are the clinical effects and brain mechanisms of open-label (honestly prescribed) placebos for chronic back pain? Findings In this randomized clinical trial of 101 adults with chronic back pain, an open-label subcutaneous placebo (saline) injection led to significant improvements in pain intensity, mood, and sleep at 1 month posttreatment compared with usual care. The placebo treatment also led to reduced somatomotor activity and increased medial prefrontal activity during evoked back pain and to increased medial prefrontal-brainstem functional connectivity during spontaneous pain. Meaning The findings of this trial suggest that open-label placebo treatments can confer meaningful clinical benefits to patients with chronic back pain by engaging prefrontal-brainstem pathways linked to pain regulation and opioidergic function.

.05.We aimed to enroll 50 per group, accounting for anticipated attrition, with n = 50 or 51 patients ultimately randomized to each group.

Randomization.
Patients were randomized to study and to treatment vs. control using an imbalance-minimization (matching) algorithm, 10 which balanced groups in number and on four covariates: pain intensity, age, gender, and opioid use (yes/no).Half of the control patients came from a simultaneous parallel study with an identical control arm (but testing a different treatment; see registered trial protocol for more details).Randomization and patient notification of group assignment was performed by YA, who had no patient contact during data collection, and group assignment was concealed from research assistants conducting data collection.
Evoked back pain.During fMRI, participants completed an evoked back pain task with a series of randomly ordered trials distending the back to one of four intensity levels.The evoked back pain task utilized a novel device providing experimental control over back pain during fMRI.Participants lay on a pneumatically-controlled cylindrical balloon, with increasing inflation causing increasingly painful back distention.The inflatable cylindrical balloon was placed under participants' lower back immediately superior to the iliac crest.Each subject received 20 trials (37 sec duration) at one of four inflation levels and rated pain after each trial on a visual analog scale (VAS; 0 = no pain, 100 = worst pain imaginable), with a total run duration of ~16 minutes (full task design details provided in eMethods p. 3).fMRI tasks design.Functional scans included an evoked back pain task, a "spontaneous pain" scan, and a thumb pressure-pain task serving as a positive control task for data quality assessment (see fMRI data quality assessment, below).
Evoked back pain task.The evoked back pain task utilized a novel device providing experimental control over back pain during fMRI.Participants lay on a pneumatically-controlled cylindrical balloon, with increasing inflation causing increasingly painful back distention.The inflatable cylindrical balloon was placed under participants' lower back immediately superior to the iliac crest.Distance from the balloon to the lateral malleolus was measured at pre-treatment and the balloon was placed was in the same location at post-treatment.Each subject received 20 trials (37 sec duration) at one of four inflation levels, and patients rated post-trial pain on a visual analog scale (VAS; 0 = no pain, 100 = worst pain imaginable).The balloon was never fully deflated during this task to limit larger head motions.The order of inflation levels for each subject was randomly permuted but constrained to optimize design efficiency by avoiding correlation with low frequency signals: Trials of inflation level 1 or 2 were always followed by inflation level 3 or 4 and vice versa, and consecutive trials always had different inflation levels.We adopted an extensive set of strategies for mitigating and controlling for head motion, described below.The evoked back pain scan was missing for one subject at post-treatment due to technical issues.Pain ratings from the evoked back pain task are shown in eFigure 2.
Spontaneous pain (resting state) scan.Participants completed a scan measuring spontaneously occurring pain (no stimulation).Participants fixated on a foveal crosshair and provided a VAS rating (7 sec) of current back pain intensity each minute.We regressed out the rating task and analyzed residual resting-state connectivity, as described below.

MRI preprocessing pipeline.
Standard fMRI preprocessing procedures were used, implemented in fMRIprep 1.2.4 11 which is based on Nipype 1.1.6. 12Anatomical T1-weighted (T1w) images from both scanning sessions were corrected for intensity non-uniformity (INU) sing N4BiasFieldCorrection 13 (ANTs 2.2.0).A T1w-reference map was computed after registration of the two T1w images (after INU-correction) using mri_robust_template. 14 The T1wreference was then skull-stripped using antsBrainExtraction.sh (ANTs 2.2.0), using OASIS as target template.Spatial normalization to the ICBM 152 Nonlinear Asymmetrical template version 2009c 15 was performed through nonlinear registration with antsRegistration, using brainextracted versions of both the T1w volume and the template.
For the functional run, first a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep.A deformation field to correct for susceptibility distortions was estimated based on two echo-planar imaging (EPI) references with opposing phase-encoding directions, using 3dQwarp 16 (AFNI 20160207).Based on the estimated susceptibility distortion, an unwarped BOLD reference was calculated for a more accurate co-registration with the anatomical reference.The BOLD reference was then coregistered to the T1w reference using flirt 17 (FSL 5.0.9) with the boundary-based registration cost-function. 18Co-registration was configured with nine degrees of freedom to account for distortions remaining in the BOLD reference.Head-motion parameters were estimated with respect to the BOLD reference before any spatiotemporal filtering using mcflirt (FSL 5.0.9).The BOLD time-series were resampled onto their original, native space by applying a single, composite transform to correct for head-motion and susceptibility distortions.The BOLD time-series were resampled to MNI152NLin2009cAsym standard space, generating a preprocessed BOLD run in MNI152NLin2009cAsym space using antsApplyTransforms, configured with Lanczos interpolation to minimize the smoothing effects of other kernels. 19noising pipeline.For both the evoked and spontaneous pain scans, nuisance covariates included 24 head motion parameters and "spike" regressors identifying volumes with framewise displacement (FD) >= .25 mm (often considered a strict threshold 20 ).Spike regression was optimized for fast-TR data by a) applying a [.1 Hz -.5 Hz] band-stop filter to head motion parameters prior to computing FD, and b) computing FD with respect to the volume collected 2.4 sec previously (5 volume difference). 21,22We additionally included spike regressors for the four volumes following an identified spike, since effects of head motion can influence subsequent volumes as well.These nuisance covariates were included in 1 st level models for evoked pain analyses and were regressed out of the spontaneous pain scans prior to connectivity analyses.
Denoising for the evoked back pain task included two additional procedures to limit the influence of head motion and remove signal less likely to be of neuronal origin.We applied anatomical CompCor, which generates nuisance covariates derived from signal fluctuations in white matter and cerebrospinal fluid, 23 and we included nuisance covariates with signal timeseries extracted from an anterior and a posterior out-of-brain area, to further capture and remove artifactual signal fluctuations in the data.
Denoising for spontaneous pain connectivity analyses additionally included global signal regression and band-pass filtering [.1 -.01 Hz], to focus on signal fluctuations most likely to be of neuronal origin. 24We also included a nuisance regressor modelling the pain rating task (boxcar regressor) in order to more closely resemble traditional resting state analyses, though prior work has found that intrinsic connectivity networks are strongly preserved during the task performance. 25del of continuous evoked pain.To limit the confounding of pain-related and ratingrelated neural signals, we collected only brief post-trial ratings (VAS, 7 sec).Since participants were in pain throughout the task, we developed an exponential decay model of continuous pain based on the post-trial ratings.
This model was validated on a separate data set collected on a subset of study participants (n = 58) who completed an evoked back pain task in a behavioral testing room during their eligibility session visit.The validation task was identical to the task administered during fMRI, except participants provided continuous pain ratings using a trackball rather than brief post-trial ratings as during fMRI.
We fit an exponential decay model to estimate continuous pain between post-trial ratings, modelling a more rapid change in pain at the beginning of each trial followed by an asymptotic approach to the next sample point.The model fit the formula f(x) = (b-a) * (1 -e^τ*x) + a, with a = pain rating at trial start, b = pain rating at trial end, x = trial timepoints between samples, and τ = a time constant governing the exponential decay process.τ was fit for each trial using the MATLAB curve fitting toolbox, and the average τ value across all subjects' trials was used to assess model performance (R 2 with bootstrapped confidence intervals, 10,000 bootstrap samples, MATLAB bootci function).
In this validation task, reported and model-predicted continuous pain were strongly related, mean R 2 = .85,95% CI = [.82 .87].Exponential decay model fits for four sample subjects are shown in eFigure 3.
Scans were excluded from analyses if pain ratings were missing on >= 25% of trials (n = 12), almost no pain was reported (pain <= 5/100 on 90% of trials, n = 3), or there was insufficient variability in pain (range <= 10/100, n = 3), as pain could not be reliably modelled in these scans.
Relationships with evoked back pain were identified by constructing a continuous withinperson estimate of evoked pain intensity based on post-trial pain ratings (continuous evoked pain model, described below).Continuous pain values were entered as a regressor in each subject's 1 st -level model (Z-scored so that voxel parameter estimates would capture pain magnitude) along with nuisance covariates, to estimate an evoked back pain parametric map for each participant at pre-treatment.fMRI data quality assessment.For the evoked back pain task, we assessed the influence of head motion at both the within-subject and between-subject level.Within-subject, we computed the variance inflation factor (VIF) for the pain regressor relative to the 24 head motion parameters, providing an estimate of task-correlated head motion.At the betweensubject level, we included a head motion summary statistic (number of volumes identified as motion outliers) as a covariate in the mixed effects model to and we tested for PRT vs. TAU differences in head motion.
For resting connectivity analyses with the spontaneous pain scans, we assessed data quality with "quality control-functional connectivity" ("QC-FC") correlations, using a whole-brain parcellation that included 489 parcels. 26,27We computed the distribution of correlations between © 2024 Ashar YK et al.JAMA Network Open.connectivity estimates and head motion across edges, an established measure of head motion associations with connectivity estimates. 20,24,28We also tested for PRT vs. TAU group differences in head motion and repeated the 2 nd level models with a head motion covariate.The median edge connectivity value was computed, and subjects' >= 3 standard deviations above the mean were excluded.
We also used a positive control task for data quality assessment.We administered 20 thumb pressure stimulations at a high and low pressure (4 and 7 kg/cm 2 ), estimated the [highlow] contrast for each subject, and applied an FDR q < .05threshold to the group average contrast map using standard general linear model analyses in SPM12.We expected significant associations with the brain responses reliably reported in acute pain tasks (e.g., midcingulate, thalamus, insula, somatosensory cortex, cerebellum).
Region of interest (ROI) definitions.ROI boundaries were defined anatomically (thalamus, PAG, RVM), from NeuroSynth (vmPFC, reverse inference map), or as spheres (16 mm radius) around peak coordinates from a previous individual patient meta-analysis of placebo analgesia. 29[31][32] Permutation testing approach.We applied a two-stage combination permutation test, as described in Winkler et al. 35 The first stage involves small volume correction within each of a series of ROIs.The second stage is a joint test of signal across ROIs to correct for the number of ROIs tested.We used a threshold of p < 0.05 Familywise Error Rate (FWER) corrected for both stages.
In Stage 1, within each ROI, we computed the exceedance mass for the observed data and compared it to a null distribution created by 5,000 permutations of the group labels (OLP vs. Usual Care).Exceedance mass was defined as the sum of the voxelwise T statistics exceeding a prespecified cluster-defining threshold.Because the choice of cluster-defining threshold is arbitrary, we confirmed that our results held across three different cluster-defining thresholds (p = .05,.01,and .005).When the exceedance mass for the observed (unpermuted) data exceeded the 95 th percentile of the null distribution (for regions hypothesized to exhibit increases) or fell below the 5 th percentile of the null distribution (for regions hypothesized to exhibit decreases), findings were considered significant.0][31][32] Use of the exceedance mass as a maximal statistic in permutation testing provides greater sensitivity than single-voxel statistics (e.g., voxelwise max T) and cluster size (number of suprathreshold voxels), as exceedance mass integrates both cluster height and extent. 33,34 Stage 2, to correct across the six ROIs tested, we constructed a combined statistic based on Fisher's method for combining p values, and we compared this statistic to its null distribution based on the permuted data.For each permutation, we computed a null-hypothesis p value for each ROI as the rank order of that permutation's exceedance mass statistic within the null distribution, and we combined those p values across the six ROIs using Fisher's formula (-2 * sum(log(p)).This provided a null distribution of the combined test statistic across ROIs. 35e observed Fisher-combined test statistic was compared to the 95% percentile of the null distribution of combined test statistics, providing a non-parametric test of whether there was more signal than expected by chance across ROIs. 35r models comparing OLP vs. UC used different covariates for each voxel (the pretreatment evoked pain response in that voxel), as this is recommended by statisticians over preto-post-intervention change scores for detecting intervention effects. 36,37Due to having different covariates for each voxel, the FSL PALM toolbox for non-parametric combination tests could not be applied for our use case.

Exploratory whole-brain results.
To conduct a sensitivity analysis investigating whether our ROI-based approach missed signal elsewhere in the brain, we conducted a wholebrain analyses thresholded at p < .005uncorrected, k = 10 mm, with results presented below.

Anatomical labeling of regions and assessment of functional connectivity profiles.
Regions were labeled using the CANLab Combined 2024 Atlas (https://sites.google.com/dartmouth.edu/canlab-brainpatterns/brain-atlases-andparcellations/2018-combined-atlas)which integrates multiple published cortical and subcortical atlases and by reference to the Julich histological atlas. 38First-and second-level models and functional connectivity profiles (see below) were estimated using SPM12 and the CanlabCore toolbox (https://github.com/canlab/CanlabCore).
To help interpret thalamic results, we visually examined the mean functional connectivity profiles of resulting clusters by estimating their functional connectivity in our data at prerandomization and quantified their probabilistic connectivity profile in the Oxford Thalamic Connectivity Atlas. 39mputation of effect sizes for clinical outcomes.We computed effect sizes (Hedges' g) of OLP vs. Usual Care group differences at each timepoint in the 1-year follow-up © 2024 Ashar YK et al.JAMA Network Open.period (change from baseline to the given timepoint, confidence intervals based on 1000 bootstrapped samples).As PGIC and Treatment Satisfaction scores were not provided at baseline, effect sizes for these measures were computed as OLP vs. Usual Care group differences at the given timepoint.
Baseline predictors of response to OLP.For archival purposes, we tested whether baseline measures of psychological functioning predicted response to OLP.We examined the following measures: pain catastrophizing (Pain Catastrophizing Scale), a one-item measure of expectations adapted from the CEQ ("At this point, how successful do you think this treatment will be in reducing your back pain?"; visual analog scale ranging from "Not at all successful" to "Very successful"), depression and anxiety (PROMIS short forms), optimism (Revised Life Orientation Test, LOT-R), and duration of chronic back pain (in years).These measures were collected at pre-randomization.For the expectation item, we asked participants to answer as if they had been randomized to treatment.
We first tested each measure as a moderator, predicting differential response to OLP vs. Usual Care.We estimated a GLM predicting post-treatment pain intensity from pre-treatment pain intensity, the measure of interest, group, and the group x measure interaction.We then tested whether each measure of interest predicted response among OLP participants only by estimating a GLM predicting post-treatment pain intensity from pre-treatment pain intensity and the measure of interest in the OLP group.

eResults. Detailed Results
fMRI data quality assessment.The positive control task produced the expected activations in pain-responsive regions (eFigure 4).Spontaneous pain scan correlations between head motion and functional connectivity estimates ("QC-FC correlations") were low, r = .02(SD = .19)across edges (eFigure 5). 20,24,28There were no group differences in head motion at preor post-treatment, both p > .55.One subject was excluded from spontaneous pain connectivity analyses due to poor data quality (median edge correlation more than 3 standard deviations above the mean).
In the evoked back pain task, subjects had M = 189.11(218.70)volumes flagged as spikes (~12% of volumes).This relatively strict approach to identifying volumes potentially corrupted by head motion still provided M = 11.03(1.67) min of data for analyses.There were no group differences in head motion at pre-or post-treatment, both p > .2. Within-subject assessment of task-correlated head motion found mean VIF = 2.9 (SD = 1.5).Only one subject had VIF >= 10, a commonly used threshold for high collinearity.There were no group differences in VIFs at either timepoint, both p > .13.Overall, this suggests a limited influence of head motion on evoked back pain estimates at both the within-and between-subject level.
Baseline predictors of response to OLP.Greater levels of pain catastrophizing at baseline predicted enhanced response to OLP.This was found in moderator analyses comparing OLP to Usual Care, group x catastrophizing β = -0.04,t(81) = -2.43,p = .017,and was marginally significant in analyses of OLP participants alone, catastrophizing β = -0.06,t(81) = -1.93,p = .06.Depression, anxiety, optimism, expectations, and duration of back pain were not significantly associated with OLP response in moderator analyses or in analyses of the OLP group alone.
Effects of OLP on BPI Severity scale.At the request of an anonymous reviewer, we conducted analyses of Brief Pain Inventory-Short Form Severity scale as an outcome.The BPI Severity scale includes 4 items, measuring last-week average, best, and worst pain, plus current pain.As described above (section Clinical Measures, p. 2) and in our preregistered analytic plan, we chose last-week average pain intensity (the first of these four items) as the primary outcome in this trial.Effects of OLP vs. Usual care on BPI Severity were in the expected direction but not significant, β = 0.33, t(91.5)= 1.38, p = 0.17.This indicates that OLP may have attenuated effects on best, worst, and/or current pain relative to last-week average pain.Further research may be needed to better understand the differential effects of treatments on average, best, worst, and current pain.
Exploratory whole-brain results.In addition to the results identified in ROI analyses, we observed OLP vs. Usual Care decreases in evoked back pain response in several areas: the right operculum, parahippocampal cortex, premotor cortex, and temporoparietal cortex (eTable 2).

. eTable 1
Effects of OLP vs. Usual Care on patient-reported outcomes at each timepoint through 1-year follow-up.

Note.
Effects of open-label placebo (OLP) vs. Usual Care on brain responses to evoked back pain.Regions were labeled using a fusion registration projection of the Julich histological atlas38 to the volumetric space of the normalization template.* indicates results significant at FWE p < .05 and depicted in Figure3; regions not indicated with an asterisk survive an exploratory p < .005uncorrected threshold and are included here for archival purposes.©2024 Ashar YK et al.JAMA Network Open.

eFigure 1 .
ROIs tested for OLP vs. Usual Care effects.We searched for OLP vs. Usual Care differences within each region (small volume corrected) and tested the joint significance of effects across regions.See text for details of ROI selection and definition.eFigure 2. Evoked back pain at pre-treatment.The left panel shows mean evoked pain at each inflation level for each patient (colored lines).The right panel shows mean evoked pain by increase in pain per increase in inflation level (β evoked pain, estimated using linear regression) for each patient.eFigure 3. Continuous pain regressors for four randomly chosen sample subjects.Grey line shows observed continuous report in the validation data, with gray circles indicating the samples taken at post-trial intervals.Predicted continuous pain between samples is shown for the linear interpolation (red) and exponential decay model (blue).eFigure4. High vs. low thumb pressure stimulation, FDR q < .05,serving as a positive control.Effects are observed in the expected pain-responsive regions.© 2024 Ashar YK et al.JAMA Network Open.
Effects of OLP vs. Usual Care on evoked back pain-related brain activity