Early effects predict trajectories of response to esketamine in treatment-resistant depression

Depression is common, disabling and often difficult to treat; after multiple treatment attempts a substantial proportion of patients remain symptomatic. Treatment-resistant depression (TRD) is commonly defined as failure to remit after at least two adequate antidepressant trials. Intranasal esketamine, the S-enantiomer of ketamine, has demonstrated rapid antidepressant effects and was approved for TRD, but its delivery requires hospital administration and two-hour monitoring, and it is costly. Given these constraints, identifying which patients will ultimately benefit from esketamine early in treatment could improve individual care and reduce unnecessary expense. Estrade and colleagues aimed to characterise distinct longitudinal trajectories of symptom change during real-world esketamine treatment for TRD and to determine whether baseline characteristics or early symptom change predict those trajectories. The study used serial Montgomery-Åsberg Depression Rating Scale (MADRS) measurements before each administration in two independent clinical cohorts; the primary objectives were to define latent classes of response and to identify predictors, including whether an early MADRS score after a small number of administrations could forecast 30-day and 90-day trajectories.

This was a longitudinal, observational study using two independent real-world cohorts of adults treated with intranasal esketamine for unipolar or bipolar TRD at two university-affiliated psychiatric departments in Paris. Patients were eligible if aged ≥18 years, met the study's definition of TRD, had a baseline MADRS ≥ 25 and were prescribed intranasal esketamine according to the standard scheme. Exclusion criteria were prior intravenous ketamine within three months, receiving fewer than three esketamine administrations, or a schizophrenia spectrum disorder. The original sample comprised 50 eligible patients (from an initial 53), and the confirmatory sample included 55 eligible patients (from an initial 67). Clinical and sociodemographic data were extracted from electronic medical records. Key variables included psychiatric and somatic comorbidities, lifetime illness characteristics (duration, number of episodes, prior ECT), current episode features (duration, Maudsley staging method scores), concomitant pharmacotherapy, and MADRS scores collected before each esketamine administration. Remission was defined as MADRS ≤ 12 and response as ≥50% reduction in MADRS at the eighth administration (approximately 28 days). Esketamine was administered following EMA-recommended phases: an induction phase (IND) of twice-weekly dosing for four weeks (initial doses 28–56 mg then typically 56 or 84 mg), followed by an optimisation/maintenance phase (OPT/MAINT) of weekly then fortnightly dosing when possible. As this was a real-world study, clinicians adjusted dosing schedules as clinically indicated. MADRS was assessed prior to every administration. For longitudinal trajectory identification, the investigators used latent class mixed models (LCMM) implemented in R (lcmm package), comparing models by Bayesian information criterion (BIC), entropy and class size, and evaluating assignment probability. Models were fit separately for the IND phase (about 30 days) and the full treatment period (about 90 days). The best-fitting model from the original sample was then applied to the confirmatory sample to test reproducibility; the association of 30-day and 90-day classifications was assessed with Somers' D. Associations between baseline characteristics and class membership were analysed with chi-square or Fisher exact tests for categorical variables and Student's t-test or Mann-Whitney U test for continuous variables, applying Bonferroni correction for multiple comparisons. Significant univariate predictors were tested in multivariate logistic regression models. To assess how well early MADRS scores predicted final trajectory, receiver operating characteristic (ROC) analyses were performed across assessments from baseline to T6, computing area under the curve (AUC) and selecting cutoffs by Youden's index; sensitivity, specificity, positive predictive value, negative predictive value and odds ratios were then calculated.

Sample characteristics: The original sample contained 50 eligible patients (mean age 53.5 years, SD 16.4) and the confirmatory sample 55 eligible patients (mean age 49.8 years, SD 17.5). Across the combined dataset, 55 patients (52.4%) met response criteria and 40 (38.1%) met remission criteria at the predefined assessment. The confirmatory cohort had greater overall severity (higher Maudsley score), more frequent suicidal ideation and a higher number of concomitant treatments compared with the original sample. Latent class analysis — original sample: In both the 30-day (IND) and 90-day (IND + OPT/MAINT) LCMMs, a two-class quadratic model adjusted for baseline MADRS provided the best fit. Class 1 showed a faster, larger and more stable decline in MADRS scores than Class 2. In the 30-day model Class 1 comprised 15 patients (30.0%) with response and remission rates of 86.7% and 73.3% respectively, compared with 40.0% and 25.7% in Class 2. In the 90-day model Class 1 comprised 35 patients (70.0%) and included 25 of the 27 responders and all 20 remitters; response and remission rates in Class 1 were 71.4% and 57.1% versus 13.3% and 0.0% in Class 2. The 30-day and 90-day classifications were significantly associated (Somers' D = 0.429; p < 0.001); 20 patients (40.0%) moved from Class 2 at 30 days to Class 1 by 90 days, and no patients moved from Class 1 to Class 2. Latent class analysis — confirmatory sample: Applying the original-sample model to the confirmatory cohort produced convergent results. In the 30-day model Class 1 comprised 28 patients (50.9%), with response and remission rates of 89.3% and 67.9% versus 11.1% and 3.7% in Class 2. In the 90-day model Class 1 comprised 31 patients (56.4%) and contained 25 of 28 responders and 18 of 20 remitters; response and remission rates in Class 1 were 80.6% and 58.1% versus 12.5% and 8.3% in Class 2. The association between the 30-day and 90-day classes was significant (Somers' D = 0.749; p < 0.001), with 7 patients (12.7%) changing class between timepoints. Baseline predictors of class membership: In the pooled 30-day model, univariate comparisons showed that Class 1 had more suicidal ideation, lower total Maudsley scores, a higher proportion of acute episode duration, fewer lifetime depressive episodes and lower rates of concomitant benzodiazepine use and lifetime ECT, while lithium use was more common. In multivariate logistic regression, a higher number of depressive episodes reduced odds of being in Class 1 (Wald χ2 = 6.2, p = 0.013; OR 0.63, 95% CI 0.44–0.90) and concomitant benzodiazepine medication also reduced odds (Wald χ2 = 5.7, p = 0.017; OR 0.24, 95% CI 0.07–0.77). For the 90-day model, drug use disorder, bipolar polarity and employment status emerged as significant predictors in multivariate analysis; the text indicates bipolar TRD increased likelihood of being in Class 1 but the extracted text does not clearly report the complete OR value for this effect. Predictive performance of early MADRS scores (ROC analyses): Across the sample, AUC values for predicting 30-day and 90-day class membership exceeded 0.5 from the first post-baseline MADRS assessment (T1). For the 30-day classes, the largest gain in accuracy occurred between T2 and T3; the MADRS score at T3 (after three administrations) predicted class membership with an accuracy of 85.7%, with a selected cutoff of MADRS ≤ 18 associated with greatly increased odds of belonging to Class 1 (OR reported but incompletely extracted). For the 90-day classes, the largest accuracy gain occurred between T1 and T2; the MADRS at T2 (after two administrations) predicted class membership with an accuracy of 80.0%. A MADRS score ≤ 22 at T2 was associated with a higher chance of belonging to Class 1, with an OR reported as 21 in the extract but without full confidence-interval detail. The authors note that a T2 MADRS ≤ 22 corresponded to a 90.9% chance of Class 1 membership, while scores >22 corresponded to a 68.0% probability of Class 2 membership.

Estrade and colleagues interpret their findings as indicating two reproducible trajectories of response to real-world intranasal esketamine in TRD: a class with fast and sustained improvement and a class with slower, less stable improvement. The two-class model replicated across independent cohorts and produced response and remission rates similar to those reported in clinical trials, though their sample included bipolar TRD cases which were generally excluded from pivotal trials. The investigators observed that some baseline factors were associated with a greater likelihood of rapid response, notably fewer lifetime depressive episodes and absence of concomitant benzodiazepine use over the first 30 days; in the 90-day model, bipolar polarity, drug use disorder and employment status were associated with a higher chance of favourable trajectory. The authors note that the benzodiazepine association concurs with some prior reports that benzodiazepines may attenuate ketamine's antidepressant effect, though other trials have not found such an effect and differences in populations and timing of assessment may explain discrepancies. A principal practical implication emphasised by the authors is that early symptomatic change strongly predicted longer-term trajectories: after two esketamine administrations the MADRS score predicted a 90-day trajectory with 80.0% accuracy, and after three administrations the 30-day trajectory could be predicted with about 85.7% accuracy. They propose that, once validated in controlled settings, such early MADRS thresholds might inform clinical decisions about whether to continue esketamine or redirect patients to alternative interventions, noting that different cutoff selection strategies (for example prioritising sensitivity over specificity) would influence false negative and false positive risks. The authors acknowledge multiple limitations: the study was observational and lacked a placebo control so effects cannot be causally attributed to esketamine; both cohorts were drawn from the same hospital system in Paris which may limit generalisability despite differences between the samples; heterogeneity and modest sample size limited power to detect less common trajectories or predictors, motivating pooling of samples; and some findings (notably differences between the 30-day and 90-day models) may be spurious and therefore require replication. Strengths cited include use of real-world data, application of latent class methods suited to identifying heterogeneous response patterns, and replication across an independent confirmatory cohort. The authors conclude that early MADRS scores and some baseline factors may help predict esketamine treatment trajectories, but they stress that optimal thresholds and causal inferences require confirmation in experimental, controlled studies.