Brain Connectivity Predicts Placebo Response across Chronic Pain Clinical Trials

Abstract

Placebo response in the clinical trial setting is poorly understood and alleged to be driven by statistical confounds, and its biological underpinnings are questioned. Here we identified and validated that clinical placebo response is predictable from resting-state functional magnetic-resonance-imaging (fMRI) brain connectivity. This also led to discovering a brain region predicting active drug response and demonstrating the adverse effect of active drug interfering with placebo analgesia. Chronic knee osteoarthritis (OA) pain patients (n = 56) underwent pretreatment brain scans in two clinical trials. Study 1 (n = 17) was a 2-wk single-blinded placebo pill trial. Study 2 (n = 39) was a 3-mo double-blinded randomized trial comparing placebo pill to duloxetine. Study 3, which was conducted in additional knee OA pain patients (n = 42), was observational. fMRI-derived brain connectivity maps in study 1 were contrasted between placebo responders and nonresponders and compared to healthy controls (n = 20). Study 2 validated the primary biomarker and identified a brain region predicting drug response. In both studies, approximately half of the participants exhibited analgesia with placebo treatment. In study 1, right midfrontal gyrus connectivity best identified placebo responders. In study 2, the same measure identified placebo responders (95% correct) and predicted the magnitude of placebo's effectiveness. By subtracting away linearly modeled placebo analgesia from duloxetine response, we uncovered in 6/19 participants a tendency of duloxetine enhancing predicted placebo response, while in another 6/19, we uncovered a tendency for duloxetine to diminish it. Moreover, the approach led to discovering that right parahippocampus gyrus connectivity predicts drug analgesia after correcting for modeled placebo-related analgesia. Our evidence is consistent with clinical placebo response having biological underpinnings and shows that the method can also reveal that active treatment in some patients diminishes modeled placebo-related analgesia. Trial Registration ClinicalTrials.gov NCT02903238 ClinicalTrials.gov NCT01558700.

Conflict of interest statement

The authors have declared that no competing interests exist.

Figures

Fig 1. Flow diagram summarizes overall experimental design, OA patients entering and completing each of the three studies, and participant subgroupings based on treatment and treatment effects.

Study 1 was analyzed to discover brain connectivity predicting placebo response. All patients received only placebo pills. Study 2 was used to validate the results from Study 1 and also to examine how the active treatment was related to placebo response. Study 2 was a double-blind randomized clinical trial. Study 3 was an observation-only trial in which no treatment was provided. Groupings and dropout causes are indicated. fMRI, functional magnetic resonance imaging.

Fig 2. Placebo response in the single-blind placebo treatment, study 1.

(A) In patients with knee OA pain (study 1), there was significant pain relief (visual analog scale [VAS], 0–10) with a 2-wk placebo treatment, which reversed to entry-level knee pain following a 2-wk placebo washout (repeated-measures analysis of variance [ANOVA], F2,32 = 6.8, p = 0.003). (B) Distribution for % analgesia (change in knee pain in VAS units from entry to 2-wk placebo treatment). The group was subdivided into placebo responders (P +) (≥20% analgesia over the 2-wk placebo treatment) and nonresponders (P −). (C) Knee OA pain shown separately for placebo responders (white) and nonresponders (black). As defined, the only decrease in pain is seen in placebo responders, after 2-wk placebo treatment. (D) Twenty knee OA pain patients (study 3), matched for age, gender, and knee VAS pain at baseline, followed over 2 wk with no treatment. There was no within-group change in knee pain over 2 wk. (E) Improvement in overall OA function (Western Ontario and McMaster Universities Osteoarthritis Index [WOMAC] scale) was observed with 2-wk placebo treatment (F1,16 = 6.21, p = 0.024). (F) The improvement was limited to placebo responders. Error bars are 95% confidence intervals (CIs). The illustrated p-values are post hoc comparisons that were statistically significant.

Fig 3. Patterns of brain connectivity in placebo responders and nonresponders in study 1.

(A) Average brain maps for degree count (number of connections of any location to the rest of the brain), shown at 10% density in placebo responders (n = 8) and non-responders (n = 9), and the difference map. Placebo responders have higher (red to yellow colors) or lower (dark to light blue) degree counts than nonresponders. (B) The brain regions highlighted were identified based on minimal t-score and threshold-free cluster enhancement (TFCE) correction. The right midfrontal gyrus (r-MFG; x = 28, y = 52, z = 9 mm; cluster 12 voxels, t-score 3.7 or p < 0.001) was the region with the highest significant between-group difference, while bilateral anterior cingulate cortex (ACC; x = −3, y = 40, z = 2; cluster 10, t-scores 2.7 or p < 0.01), posterior cingulate cortex (PCC; x = −1, y = −45, z = 15; cluster 14, t-score −2.7 or p < 0.01), and a right region overlapping the secondary somatosensory and primary motor cortex (r-S2/M1; x = 60, y = −7, z = 21; cluster 31, t-score of 1.7 or p < 0.05) had lower significant differences. (C) Degree counts derived from r-MFG region in OA patients classified as placebo responders and nonresponders, and in healthy subjects (n = 20), for densities 2%–20%. At all densities, placebo responders (white) show higher degree counts than placebo nonresponders (black, group * density F9,135 = 15.3, p < 0.0001) or healthy controls (green, F9,234 = 5.8, p < 0.0001). (D, E) r-MFG degree counts at 10% density significantly predicted future (2-wk) magnitude of % analgesia for all OA patients based on both VAS and WOMAC scores. In C–E; black, gray, and white symbols represent the same groups as in Fig 1, while green symbols represent the healthy controls on which brain imaging was performed.

Fig 4. Pain relief in the double-blind placebo-controlled 3-mo duloxetine treatment, study 2.

(A) Of all the knee OA pain patients who participated in study 2, 20 were randomized to placebo (P, grey) and 19 to duloxetine (DLX, red) treatment. Both groups started at the same level of knee pain (VAS) and exhibited significant and similar magnitudes of pain relief with a 3-mo treatment (only time-effect repeated-measures ANOVA, F1,37 = 14.8, p < 0.0001). (B) Participants in both arms were classified as responders (P +, white; DLX +, pink) or nonresponders (P −, black; DLX −, red) (≥20% analgesia over the 3-mo placebo treatment). Both treatments resulted in similar numbers of improvers and similar magnitudes of pain relief, observed, by design, only in the treatment responders (white and pink). (C) Twenty knee OA patients (study 3), matched for age, gender, and knee VAS pain at baseline, followed over 3 mo with no treatment (green). There was no within-group change in knee pain over 3 mo of no treatment. (D, E) We observe the same pattern of symptom relief, as observed for VAS, when the WOMAC scale is used as an outcome measure (only time-effect repeated-measures ANOVA, F1,37 = 13.3, p = 0.001). Error bars are 95% CIs. The illustrated p-values are post hoc comparisons that were statistically significant.

Fig 5. Predicting placebo and duloxetine treatment outcomes from r-MFG degree counts in study 2.

Prediction of future outcomes (A, B for placebo treatment; C, D for duloxetine treatment) was assessed for r-MFG degree counts (based on brain coordinates derived from study 1). (A) r-MFG degree counts were significantly higher in placebo responders (post hoc honestly significant difference [HSD] test, p = 0.001), and the receiver operating characteristic (ROC) curve identified grouping at 95% accuracy. (B) Empirical analgesia was correlated to analgesia predicted from the best-fit line calculated in study 1, using r-MFG degree counts from study 2, for VAS (p = 0.004) and more weakly for WOMAC (p = 0.12) outcomes. (C) In contrast, in patients randomized to duloxetine, the r-MFG degree count did not differentiate between responders and nonresponders (t-score17 = 1.5, p = 0.17; ROC area under the curve [AUC] = 0.67) and (D) did not predict empirical analgesia for VAS and WOMAC outcomes. Error bars are 95% CIs. Symbol colors represent the same groups as in Fig 3.

Fig 6. Right parahippocampal gyrus (r-PHG) degree counts predict future duloxetine response, based on modeling the placebo response in duloxetine-treated patients in study 2.

(A) The empirical analgesia of individual duloxetine-treated patients (red) and the predicted placebo response (grey) are illustrated. The predicted placebo response was derived from the best-fit equation from study 1, which was applied to r-MFG degree count in duloxetine-treated patients. Patients with minimal predicted placebo and ≥20% empirical analgesia were considered mostly duloxetine responders (subjects 4 and 6; arrows). (B) Contrasting the whole-brain degree counts of these two subjects with the six other duloxetine responders (subjects 1, 2, 3, 5, 7, and 8) revealed a right parahippocampal gyrus region (r-PHG) in which degree counts were higher in subjects 4 and 6 (scatter of individual values and median and quartiles are shown; Mann-Whitney U-test, p = 0.071). (C) r-PHG degree count correlated with the difference between empirical analgesia and predicted placebo response for VAS (p = 0.048) and WOMAC (p = 0.033) outcomes, suggesting that the regional functional connections also reflect future placebo-corrected drug response for all 20 duloxetine-treated patients.