International pooled patient-level meta-analysis of ketamine infusion for depression: In search of clinical moderators

Ketamine, a glutamatergic anaesthetic, has shown rapid and large antidepressant effects at subanaesthetic intravenous doses in randomized controlled trials, including in treatment-resistant and bipolar depression. However, most ketamine RCTs have small samples that are adequate to detect group-level effects but underpowered to identify moderators—baseline characteristics that predict which patients will benefit most. Previous candidate predictors (clinical, mechanistic, biological) have not been consistently replicated across trials, and study-level meta-analyses have failed to find reliable moderators. This uncertainty limits personalised prescribing and the efficient targeting of ketamine to patients most likely to gain clinically meaningful benefit. Price and colleagues therefore conducted an individual participant data (pooled patient-level or "mega-analytic") meta-analysis of randomized IV ketamine trials recruiting patients with depressive symptoms. The study aimed to (1) quantify ketamine's effect versus control on continuous and dichotomous depression outcomes at two acute timepoints (~24 hours and ~7 days after a single infusion); (2) test patient- and study-level moderators of ketamine's efficacy using sequential (univariate) analyses; and (3) apply a data-driven combined-moderator method to identify weighted combinations of variables that might improve clinical prediction. The focus was on moderators readily available in clinical practice and on trials using clinician-rated MADRS or HRSD-17 depression scales.

Trials were identified through a pre-registered systematic search of PubMed to 19 January 2021 and supplemental checks of published reviews. Eligible studies were randomized controlled trials in which at least one IV ketamine infusion was administered to adults with unipolar or bipolar depression or another disorder where depressive symptoms are central (for example, PTSD). Trials were excluded if ketamine was given alongside other study-administered treatments such as ECT, but concomitant clinically prescribed medications were allowed. Authors of eligible trials were invited to contribute individual participant data, including drug condition, infusion order for crossover trials, pre- and post-infusion MADRS and HRSD-17 scores, and a prespecified set of 33 potential moderator variables. Target timepoints were "rapid" (≈24 hours post-infusion, allowable window 4 hours to 3 days) and "post-rapid" (≈7 days post-infusion, allowable window 6–14 days). Contributing teams attested to trial methodological details used to assess risk of bias across Cochrane-relevant criteria. Risk of bias was judged low overall except for possible functional unblinding from ketamine-specific side effects. To harmonise outcomes, HRSD-17 scores were converted to MADRS using a published conversion algorithm when necessary, and analyses focused on patients receiving IV ketamine ≥0.5 mg/kg versus placebo (inert or psychoactive). For trials with crossover or repeated-infusion designs, only data from the first infusion were used to avoid within-subject repeated measures complicating moderator tests. Availability of the 33 moderators varied widely across participants and studies. Analyses used linear mixed-effects regression models with a random effect for study to account for between-study heterogeneity; continuous moderators were standardised and dichotomous variables coded 0/1. Completer datasets were used because completion rates were high (≥90%) and imputation across studies was deemed of limited value. Primary outcomes were percentage improvement in MADRS from baseline to rapid and post-rapid timepoints. Response (≥50% MADRS reduction) and remission (MADRS ≤ 9) were reported descriptively but not used in moderator testing. Moderators were examined sequentially (interaction moderator*treatment), with a pre-specified set of nine "Tier 1" variables that were non-redundant and available in ≥99.5% of patients treated as primary; Bonferroni correction was applied across these nine tests. An additional 29 "Tier 2" variables available in ≥40% of patients were treated as exploratory and not adjusted for multiple comparisons. Moderator effect sizes were expressed as Spearman r correlations; the authors considered |r| ≥ 0.3 to be of potential clinical utility. A data-driven combined moderator (M*) approach was also applied. This uses regularised multivariable regression to derive weights for multiple moderators, producing a single composite score that may yield stronger predictive power than individual moderators. Separate M* models were run for rapid and post-rapid outcomes. Tier 1 patient-characteristic variables were included in M* #1; Tier 2 variables were grouped thematically into seven subsets (M* #2a–2g) to maximise sample size within each combined analysis. Bootstrap confidence intervals were used to assess statistical significance of M* estimates.

Seventeen of 25 eligible trials provided at least one usable outcome variable, yielding pooled data from n = 809 patients; n = 720 patients received one of the ketamine or control conditions specified for inclusion in the primary meta-analyses. After harmonisation, moderator availability across the requested 33 variables ranged from 10.8% to 100% (median 55.6%). Ketamine dosing and administration were largely uniform across included trials, supporting the primary exposure definition of IV ketamine ≥0.5 mg/kg. Main clinical outcomes confirmed robust antidepressant effects at both timepoints. The pooled analyses showed strong reductions in depression severity at the rapid (≈1 day) and post-rapid (≈7 days) assessments, with overall responder and remission peaks reported as 46% and 27% respectively across the pooled sample. These rates were comparable to those reported in clinical settings but lower than earliest RCTs in the literature. Sequential moderator analyses tested 37 potential moderators. Three significant Tier 1 moderators emerged, all pertaining to study-level design features. First, trials that required a higher treatment-resistant depression (TRD) threshold for enrolment (≥2 failed adequate antidepressant medication trials) showed a larger ketamine effect relative to placebo at the post-rapid timepoint; this interaction survived correction for multiple comparisons (post-rapid r = 0.108; β* 0.47, 95% CI 0.16 to 0.77; p adjusted = 0.027). The pattern reflected somewhat larger ketamine responses and somewhat lower placebo responses in higher-TRD trials. Second, crossover-design trials displayed greater ketamine-versus-placebo separation at the rapid timepoint (r = 0.132; β* 0.52, 95% CI 0.23 to 0.81; p adjusted = 0.036), an effect driven mainly by lower placebo responses in crossover studies rather than larger ketamine responses. A third study-level moderator—trials conducted in the US—showed a larger post-rapid ketamine effect unadjusted (r = 0.089; β* 0.41; p unadjusted = 0.0096) but this did not survive correction. Among Tier 2 (exploratory) moderators, higher baseline systolic blood pressure predicted greater post-rapid benefit from ketamine relative to placebo (r = 0.106; β* 0.23, 95% CI 0.04 to 0.42; p unadjusted = 0.019). Several other variables (placebo type, marital status, Black race, number of failed trials at patient level, number of major depressive episodes, BMI) showed trend-level moderation (p unadjusted < 0.10) in at least one analysis but were not robust. Combined moderator (M*) analyses found that each composite model was statistically significant (bootstrap CIs did not cross 0), and all M* effect sizes exceeded the largest single-moderator effect. Nevertheless, composite effect sizes remained small-to-medium, with r values ranging from 0.12 to 0.29 across models. The largest differential effect was observed in M* #2f, which combined six Tier 1 variables with BMI and smoking status in a dataset of n = 232 patients from seven studies. For the rapid timepoint this model produced r = 0.293 (95% CI 0.175 to 0.415) and for the post-rapid timepoint r = 0.234 (95% CI 0.118 to 0.347). Variables with the largest weights in this combined moderator were study TRD threshold, current MDD diagnosis, study country (US vs outside US), and BMI; the pattern associated with greater ketamine-versus-placebo improvement included greater prior treatment resistance, absence of an MDD diagnosis (for example bipolar disorder or PTSD), US-based study, and higher BMI. Despite statistical significance, the authors characterised these combined effects as modest and of limited clinical utility because a large portion of variance in outcome remained unexplained.

Price and colleagues interpret the findings as confirmation that a single IV ketamine infusion produces robust, rapid antidepressant effects across a heterogeneous international sample, with clinically meaningful response and remission occurring in a substantial minority of patients. The pooled patient-level approach increased power for moderator testing but nevertheless identified few reliable patient-level clinical or demographic moderators that could usefully guide individualised treatment selection prior to a first infusion. When moderators were detected, they were mainly study-design features: higher TRD thresholds at enrolment and crossover designs were associated with greater ketamine-versus-placebo separation, and an apparent US study effect was observed but did not survive correction. The authors position these results in relation to prior trial-level work by noting that many previously reported candidate moderators (for example age, sex, diagnosis of unipolar versus bipolar depression, concomitant medication status) did not predict differential benefit in this pooled dataset. They suggest that study design choices—particularly TRD inclusion thresholds—can influence observed effect sizes and that reserving ketamine for patients with greater prior treatment resistance may both align with consensus clinical recommendations and enhance trial power. The enhanced separation in crossover trials may reflect expectancy effects or other design-specific influences, given that only first-infusion data were analysed and carry-over was therefore unlikely to explain the finding. Key limitations acknowledged by the investigators include the predominance of single-infusion RCTs in the literature (limiting inferences about serial infusion regimens used in clinical practice), lack of longer-term follow-up data, and a constrained set of harmonisable moderators across trials. Some moderators could only be examined at the between-study level, reducing power and limiting the advantages of the pooled participant-level approach. The dataset also lacked diversity in age ranges (no paediatric or geriatric trials met inclusion criteria at the time), had limited racial and ethnic heterogeneity, and under-represented patients with common comorbidities seen in routine clinical settings. The authors note that intranasal esketamine trials were not included owing to fewer published studies and data-sharing constraints, which may limit generalisability to that licensed treatment. In terms of implications, the authors conclude that, because routinely collected clinical and demographic variables have limited predictive power, a pragmatic "fast-fail" clinical strategy—testing a short, time-limited course of one to three infusions to identify responders—remains the most accurate available approach to individualise ketamine treatment. They further suggest that development and integration of mechanistic, treatment-relevant biomarkers (for example neuroimaging or accessible blood-based measures) may be necessary to achieve clinically useful precision in predicting ketamine response, although such measures are currently less available and insufficiently replicated.

This international pooled patient-level meta-analysis validated the rapid and short-term antidepressant efficacy of a single IV ketamine infusion, with overall responder and remission rates of approximately 46% and 27% respectively. Despite exhaustive sequential and data-driven combined moderator searches across 37 variables, the ability to predict which patients will preferentially benefit from ketamine prior to treatment was limited. The investigators therefore recommend that, until more predictive mechanistic markers are established, a time-limited trial of ketamine (a "fast-fail" strategy) remains the most accurate method to identify individual responders, while noting access, cost, and equity issues that may impede this approach in many healthcare contexts.