Neuroimaging correlates and predictors of response to repeated-dose intravenous ketamine in PTSD: preliminary evidence

Norbury and colleagues situate the study in emerging evidence that intravenous ketamine can reduce symptoms of post-traumatic stress disorder (PTSD) beyond its antidepressant effects. They note that ketamine may promote synaptogenesis and a transient state of increased neuroplasticity in medial prefrontal and hippocampal circuits, which could facilitate un- or re-learning processes such as extinction of fear memories. Prior imaging work in depression and PTSD has implicated altered prefrontal–amygdala circuitry during emotional processing, but neural correlates of symptomatic change after ketamine treatment in PTSD had not been reported. This paper reports a preliminary, hypothesis-driven analysis of functional MRI data collected before and after repeated-dose intravenous ketamine or midazolam in a subset of participants from a randomised clinical trial (N = 21). Using an a priori target circuit (vmPFC, rACC, dACC, anterior insula, anterior hippocampus, amygdala) and pre-registered analytic procedures, the investigators sought to identify imaging measures that correlate with or predict PTSD symptom change. Exploratory follow-up analyses, including dynamic causal modelling, were used to probe symptom specificity and the directionality of connectivity effects.

The imaging substudy drew participants from a randomised clinical trial of six intravenous infusions of ketamine (0.5 mg/kg) versus psychoactive placebo midazolam (0.045 mg/kg), administered three times per week for two weeks. Participants and study staff were blinded to allocation. Twenty-one trial participants meeting DSM-5 criteria for chronic PTSD consented to pre- and post-treatment MRI; post-treatment scans were taken after at least one week of treatment (timing varied for some participants). Primary clinical measures were the Clinician-Administered PTSD Scale for DSM-5 (CAPS-5) and the Montgomery–Åsberg Depression Rating Scale (MADRS), both administered by blinded raters. Side-effect measures (dissociative, psychotomimetic, manic, somatic) and medication/toxicology status were recorded. Each imaging session included a T1-weighted structural scan and three T2*-weighted functional runs: (1) an emotional face-processing task (block design faces vs shapes); (2) an emotional conflict regulation task (face Stroop; incongruent vs congruent trials); and (3) a 12-minute eyes-open task-free resting-state scan. Data were acquired on a 3 T Siemens scanner and pre-processed with fMRIprep; first-level models were specified in SPM12. Primary contrasts were faces>shapes and incongruent>congruent. To limit motion confounds, all imaging measures were regressed against mean framewise displacement and residuals used in subsequent analyses. Analyses focused on a predefined target circuit informed by prior PTSD imaging literature. Region-of-interest masks were derived from anatomical atlases and meta-analytic maps. Extracted imaging measures included mean regional BOLD responses, ROI-to-ROI functional connectivity via generalized psychophysiological interaction (gPPI), and multivariate representational similarity metrics (linear discriminant contrasts). The primary pre-registered statistical approach used elastic net penalised regression with leave-one-subject-out cross-validation (LOOCV) to identify imaging features related to change in CAPS-5 scores and to identify baseline predictors of symptom change. Elastic net balances L1 (feature selection) and L2 (shrinkage) penalties to handle situations with more predictors than observations and correlated predictors; model hyperparameters were chosen to minimise mean squared error in left-out subjects. Models were run both drug-agnostic and including drug-by-imaging interaction terms; a further model also adjusted for change in MADRS to assess PTSD-specific effects. Covariates included demographic and clinical variables and side-effect measures. Missing-data handling, sample-size details, and deviations from preregistration are reported in the Supplementary Material. Exploratory follow-up analyses examined symptom-dimension specificity using PTSD subscales and probed effective connectivity (directionality) using dynamic causal modelling (DCM). DCM model structure for the vmPFC–amygdala pair was informed by prior studies; hierarchical Parametric Empirical Bayes (PEB) was used to relate changes in effective connectivity to symptom change and drug interactions.

The imaging subsample (N = 21) comprised individuals with severe, chronic PTSD (mean baseline CAPS-5 = 37; mean illness duration 17 years). Eleven participants received ketamine and ten midazolam. Both groups showed symptom improvement, but the ketamine group improved more strongly on CAPS-5 and MADRS (significant session*drug interaction in the parent trial and the imaging subsample). Pre-registered elastic net models relating change in imaging measures to change in PTSD severity identified increased functional connectivity between the ventromedial prefrontal cortex (vmPFC) and amygdala during emotional face-viewing as the strongest correlate of symptom improvement across all participants (standardized β = 2.90). Additional retained correlates were increased rACC BOLD during negative emotional conflict regulation (β = 0.97) and greater resting rACC–anterior insula connectivity (β = 0.68); the model explained ~32% of variance (whole-model r2 = 0.318). When drug-by-imaging interactions were included, the vmPFC–amygdala change effect interacted with drug, indicating a stronger relationship in ketamine-treated individuals (interaction β = 0.86). In this interaction model, decreases in dACC BOLD during emotional conflict regulation and increased resting vmPFC–anterior insula connectivity were also retained as predictors, with the dACC effect largely present only in the ketamine group (interaction β = -2.82, vmPFC–anterior insula interaction β = 0.60; whole-model r2 = 0.585). Including concurrent change in MADRS did not abolish the vmPFC–amygdala effect (β = 0.69) and retained the drug interaction for dACC, suggesting some specificity to core PTSD symptoms (model r2 = 0.667). Exploratory symptom-dimension analyses showed that increased vmPFC–amygdala coherence during face-viewing related most strongly to improvements in re-experiencing, avoidance, and anxious arousal (rs ≈ 0.53–0.56) and less strongly to negative affect or anhedonia (rs ≈ 0.19–0.45). Change in vmPFC–amygdala connectivity was not related to number of infusions at the time of the post-scan or days between scans. Dynamic causal modelling (DCM) and PEB analyses examined directionality. At baseline, emotional faces tended to modulate the amygdala→vmPFC pathway excitatorily (Ep = 0.89, posterior probability Pp = 1.00) with weak evidence for vmPFC→amygdala inhibition. Across pre-to-post changes, greater PTSD improvement (all subjects) was associated with reduced face-related excitation from amygdala→vmPFC (Ep = -0.022, Pp = 0.98) and increased face-related vmPFC→amygdala inhibition (Ep = -0.052, Pp = 1.00). Separated by drug, the reduction in amygdala→vmPFC excitation related to improvement in both midazolam and ketamine groups, but increased vmPFC→amygdala inhibition was evident only in ketamine responders (ketamine Ep = -0.091, Pp = 1.00; midazolam Ep = 0.004, Pp = 0.60). A 3rd-level PEB comparison supported a stronger vmPFC→amygdala inhibition effect in the ketamine group (Ep = -0.047, Pp = 0.98). Baseline prediction models identified lower pre-treatment vmPFC–amygdala connectivity during face-viewing as the strongest predictor of greater PTSD symptom improvement across all participants (β = -6.03). Lower baseline rACC BOLD during emotional tasks and greater representational differentiation of fearful vs neutral faces in rACC voxels were also predictive (βs = -4.83, -1.30, and 1.35 for the respective measures). Small effects for baseline clinical severity and concurrent psychotropic or cannabis use were retained but may reflect regression to the mean. When drug interactions were modelled, the relationship between low baseline vmPFC–amygdala connectivity and improvement was stronger in those who later received ketamine (interaction β = -4.15; model r2 ≈ 0.865). Baseline DCM suggested that poorer vmPFC inhibition of the amygdala and stronger amygdala→vmPFC excitation during face viewing predicted greater subsequent improvement, particularly for ketamine-treated participants. The authors note the high whole-model r2 in some models but caution these may reflect over-fitting. Additional notes on task performance: accuracy and reaction times on the emotion tasks were near ceiling and did not vary by session or drug. The authors also regressed imaging metrics on mean framewise displacement to reduce motion confounds and used imputation where needed; they report nearly 20% imputed data for the face Stroop task.

Norbury and colleagues interpret their findings as preliminary evidence that improvement in PTSD following repeated-dose intravenous ketamine (and to a lesser extent midazolam) is accompanied by changes in fronto-limbic circuitry, particularly increased vmPFC–amygdala coherence during processing of socio-emotional stimuli. They highlight that this pattern fits with prior literature showing hypoactive prefrontal and hyperactive amygdala responses in PTSD, and with observations that increased prefrontal–amygdala connectivity is associated with successful exposure-based treatment. Directionality analyses suggested a two-part picture: across both drugs, clinical improvement was associated with decreased excitation from the amygdala to the vmPFC during face viewing, which the investigators characterise as a possible common anxiolytic pathway. However, increased top-down inhibition from vmPFC to amygdala during emotional stimuli was observed only in ketamine responders, a drug-specific effect confirmed in higher-level PEB comparisons. The authors propose, cautiously, that this ketamine-specific normalisation of prefrontal control over amygdala responses could reflect restoration of circuitry important for reducing threat responsivity and promoting extinction learning. They link this to the hypothesis that ketamine opens a transient window of neuroplasticity in medial prefrontal cortex that might facilitate un- or re-learning of maladaptive fear associations. The discussion emphasises that baseline hypofrontal control of the amygdala (lower vmPFC→amygdala inhibition) predicted greater clinical benefit from ketamine, suggesting that individuals with this circuitry profile may be more likely to respond. Other baseline markers associated with response included lower rACC activation and greater representational differentiation of fearful faces in the rACC, measures the authors suggest may index a PTSD phenotype marked by altered threat responsivity and extinction-related circuitry. Key limitations acknowledged by the authors are the small sample size and consequent limited power to detect small-to-moderate effects, the risk of inflated effect sizes in small samples, and the possibility of over-fitting despite cross-validation. Measurement reliability concerns for univariate BOLD measures—especially in subcortical regions—are also noted, along with the relatively high proportion (~20%) of imputed data for the face Stroop task. The authors therefore present the results as hypothesis-generating rather than definitive. Implications discussed include the potential relevance of combining plasticity-promoting pharmacological treatments such as ketamine with psychological interventions that directly target extinction or reconsolidation processes, to test whether such combinations improve response rates or prolong benefit. The authors call for replication in larger, independent samples and for studies designed to test causality and mechanisms more directly.