Published on 20 September 2026
rTMS or esketamine for treatment-resistant depression: a target trial emulation that still can’t settle the choice
Journal of Affective Disorders · 2026 · volume 399 · article 121132 · issue of April 15, 2026 · Chen et al.
DOI 10.1016/j.jad.2025.121132
PMID 41506390
Scientific 60
Editorial 69
The essentials
Direct comparative data between repetitive transcranial magnetic stimulation and esketamine in treatment-resistant depression remain scarce, and this target trial emulation on US electronic health records tries to fill that gap: 1,802 patients starting stimulation and 7,465 starting esketamine before matching, then 1,690 patients in each arm after 1:1 propensity-score matching on 50 baseline characteristics, followed from day 14 to day 365. On suicidal outcomes, no difference is shown over the whole follow-up, with a hazard ratio of 1.199 and a 95% confidence interval of 0.943 to 1.526. On three tolerability outcomes, hospitalization or emergency visits, arrhythmia and injuries, the values favor the stimulation arm. Two objections limit heavily what can be drawn from this, and the authors themselves state part of it. Depressive severity was neither collected nor matched, although the two initial populations differed markedly on their propensity score. And esketamine is administered under medical supervision after every dose, while outcomes are built from codes generated at medical encounters, which mechanically creates more opportunities to record an event in that arm: the authors explicitly list this among their limitations. Finally, the word efficacy here covers coded suicidal outcomes and care-seeking that the authors themselves call indirect indicators: no symptom scale, no response rate, no remission rate appears in the publication.
The context
This question comes up several times a month in consultation. A patient has failed two lines of antidepressants, both options are available, and nothing says which to offer first. Without a robust direct comparison, clinicians have had to weigh separate trials against each other, run on different populations with different criteria. The authors note that both treatments hold approval from the US drug agency for this indication. The French regulatory status is not addressed by the publication and is therefore not discussed here.
A target trial emulation means writing the protocol of the randomized trial one would like to run, then reproducing it as faithfully as possible on already collected data. It is a serious framework, now common in pharmacoepidemiology. It corrects several classic errors of observational reasoning, and the authors state that they pre-specified every element of the protocol before analysis, including time zero, the follow-up window and the statistical plan, synchronizing eligibility, treatment assignment and the start of follow-up to limit immortal time bias. It does not create randomization, and that is where the reading of this study turns. This judgment on the scope of the methodological framework is expert opinion, distinct from the reported results.
The study at a glance
| Question (PICO) | |
|---|---|
| Population | |
| Adults aged 18 to 65 with treatment-resistant depression, defined here as prior exposure to two distinct antidepressants. Diagnosis of a depressive episode, recurrent depressive disorder or dysthymic disorder (F32, F33, F34.1), with at least three recorded encounters. Excluded for bipolar disorder (F31), prior exposure to the other group’s treatment, or prior exposure to the same treatment before the index event. Out of 2,051,453 candidates identified, 1,802 patients under stimulation and 7,465 under esketamine were retained before matching. | |
| Intervention | |
| Repetitive transcranial magnetic stimulation, identified by CPT codes 90867, 90868, 90869 and 1020574, 1,690 patients after matching. | |
| Comparator | |
| Esketamine, identified by drug dispensing (RxNorm 2119365), 1,690 patients after matching. | |
| Outcomes | |
| Seven outcomes pre-specified in the methods: suicidal outcomes (X71 to X83, T14.91, R45.851), emergency visits or hospitalizations of any cause, hypertension (I10 to I1A), arrhythmia (I47, I48, I49), seizure (G40, R56), injuries of any cause (S00 to S99, T07, T14), death. Follow-up from day 14 to day 365, split into four pre-specified windows: 14 to 30, 30 to 90, 90 to 180 and 180 to 365 days. Subgroup analyses by age (18 to 44 versus 45 to 65) and by sex, plus sensitivity analyses. | |
| Design | |
| Retrospective cohort study on de-identified electronic health records from the TriNetX US Collaborative Network, 72 healthcare organizations, with target trial emulation and 1:1 propensity-score matching on 50 baseline characteristics, balance assessed by a standardized mean difference below 0.10. Treatments initiated between March 5, 2019, the date of US approval of intranasal esketamine for this indication, and September 10, 2025. Cox models, Kaplan-Meier curves and log-rank tests. Protocol approved by the Kaohsiung Medical University Hospital ethics committee, consent waived. |
Quality control
| Point checked | Judgment |
|---|---|
| Methodological construction | Pre-specified |
| FindingA target trial emulation is claimed and described element by element, with 1:1 propensity-score matching on 50 baseline characteristics, a STROBE checklist provided, and the queries defining both cohorts published line by line with their terminologies and codes. The supplementary figure shows propensity-score distributions that diverge clearly before matching and overlap after. The authors state they synchronized eligibility, assignment and the start of follow-up to limit immortal time bias. The choice to start follow-up at day 14 is not justified in the text. | |
| Depression severity measurement | Not collected |
| FindingThe authors acknowledge the point themselves: depressive severity was recorded neither in the analysis nor in the initial matching, since standardized scales are not consistently available in TriNetX. They report substituting indirect indicators, routine clinical history and care-seeking before the index event, and specify that finer indicators, such as the exact number of prior antidepressants, could not be reliably operationalized. Matching balances only what is measured. | |
| Nature of outcome measures | Acknowledged bias |
| FindingEvery outcome is a code produced at a medical encounter, while intranasal esketamine is administered in the presence of a clinician and requires monitoring after every dose. This point concerns the product’s conditions of use rather than the trial itself. The authors list it among their limitations: differences in care organization and monitoring intensity can increase contact with the healthcare system and inflate the number of recorded emergency visits and hospitalizations under esketamine. | |
| What “efficacy” measures | Indirect indicators |
| FindingThe abstract announces a comparison of efficacy and tolerability. Efficacy is here operationalized through suicidal outcomes and care-seeking. The authors specify that the psychiatric nature of hospitalizations and emergency visits could not be confirmed, and that these outcomes should be read as indirect indicators, not as direct measures of antidepressant effect. No symptom score, response rate or remission rate appears in the publication. | |
| Completeness of results reporting | Gaps |
| FindingNo event counts and no absolute risks are reported: results are given only as hazard ratios, log-rank tests and survival curves. Death is listed among the outcomes announced in the methods and in the flow diagram, but no mortality result is presented. Conversely, delirium appears in the results and in the sensitivity analysis without having been defined among the pre-specified outcomes. Testing of the proportional hazards assumption is announced through generalized Schoenfeld residuals, but its result is not reported, even as hazard ratios vary widely from one follow-up window to another. | |
| Multiplicity of comparisons | Uncorrected |
| FindingSeven outcomes pre-specified, six analyzed over the whole follow-up and then across four windows, complemented by age and sex subgroup analyses and two sensitivity analyses. No adjustment for multiple comparisons is mentioned, the retained threshold being a two-tailed p below 0.05. The authors themselves ask, twice, that window-level and subgroup results be interpreted with caution because of these multiple comparisons and the low number of events for some outcomes. | |
| Priority claim | Caveat stated |
| FindingThe authors write, in the discussion, that this is to their knowledge the first and largest real-world study using a target trial emulation framework to compare magnetic stimulation and esketamine in treatment-resistant depression. The restriction is explicit and holds: the prior comparisons they cite themselves concern ketamine rather than esketamine, a retrospective naturalistic study, and a post hoc pooled analysis of two randomized trials. | |
| Funding and conflicts of interest | No ties declared |
| FindingThe publication states that this work received no specific grant from public, commercial or non-profit bodies, and that all authors declare no conflict of interest. The authors further report limited use of generative artificial intelligence tools for language polishing and figure code, specifying that no analysis or result was produced by these tools. The STROBE checklist marks the funding item as “not applicable,” which is inconsistent with the declaration present in the article. | |
| Reproducibility | Partial |
| FindingA genuine strength: the queries defining both cohorts are published in full, terminologies and codes included, which makes patient selection replayable, and the sensitivity analyses include a negative control. Weaknesses: the data platform is proprietary and access-controlled, the full list of the 50 matched characteristics is not published, the baseline characteristics table presenting, by the authors’ own admission, only the main ones for lack of space, and no reference to a public pre-analysis protocol registration appears in the publication. A minor discrepancy remains between the flow diagram, which mentions excluding patients with fewer than two encounters, and the text and supplementary queries, which require at least three encounters. | |
The results
| Result | Value |
|---|---|
| Suicidal outcomes, whole follow-up | 1.199, interval 0.943 to 1.526 |
| ReadingNot significant, log-rank p equal to 0.118. The interval remains compatible with a difference in either direction. An undemonstrated difference is not a demonstration of equivalence, and this symmetry holds for both treatments. | |
| Hospitalization or emergency visit | 0.746, interval 0.646 to 0.861 |
| ReadingFewer events recorded in the stimulation arm, log-rank p below 0.001. This is an all-cause composite outcome whose psychiatric origin could not be confirmed, and it is directly exposed to detection bias linked to the difference in monitoring. | |
| Arrhythmia | 0.677, interval 0.485 to 0.943 |
| ReadingSame caveat. A rhythm disorder is more likely to be coded when the patient is examined after every administration. | |
| Injuries, all causes | 0.745, interval 0.627 to 0.885 |
| ReadingSame caveat. The gap is largest in the 14 to 30 day window, at 0.383 with an interval of 0.257 to 0.570, then narrows to become non-significant between 180 and 365 days, at 0.873 with an interval of 0.683 to 1.115. | |
| Hypertension | Not significant over the whole follow-up |
| ReadingLog-rank p equal to 0.115, with no overall hazard ratio reported. One significant difference only, in the 14 to 30 day window, at 0.661 with an interval of 0.488 to 0.896, favoring stimulation. The following windows show no further difference. | |
| Seizure | Not significant over the whole follow-up |
| ReadingLog-rank p equal to 0.057, with no overall hazard ratio reported. Window-level estimates are all below unity, but none reaches the threshold, and the 14 to 30 day window is not estimable, reflecting a very low number of events. | |
| Suicidal outcomes, 30 to 90 day window | 1.496, interval 1.031 to 2.170 |
| ReadingThe only significant window on this outcome, favoring esketamine. The three other windows point the same way without reaching the threshold: 1.371 at 14 to 30 days, 1.210 at 90 to 180 days, 1.089 at 180 to 365 days. The lower bound sits very close to unity, within a set of uncorrected multiple comparisons. The authors themselves ask that this result be interpreted with caution. | |
| Age and sex subgroups | 1.558 between ages 45 and 65, interval 1.027 to 2.365 |
| ReadingThe excess of suicidal outcomes under stimulation is significant only among patients aged 45 to 65, the value standing at 1.116 with an interval of 0.820 to 1.517 among those aged 18 to 44. Among men, stimulation’s tolerability advantage is more pronounced: 0.629 for hospitalization or emergency visits, 0.479 for arrhythmia, 0.616 for injuries. Uncorrected subgroup analyses, to be treated as hypothesis-generating. | |
| Sensitivity analyses | Negative control 0.730, interval 0.462 to 1.153 |
| ReadingThe chosen negative control, infectious cellulitis, shows no difference between groups, which argues against a generalized detection bias without ruling it out for cardiovascular outcomes and care-seeking. Restricting the index period to 2022 through 2025, the tolerability differences persist, but the hazard ratio for suicidal outcomes moves from 1.199 to 0.901 with an interval of 0.664 to 1.225, that is, it changes direction while remaining non-significant: the central estimate for this outcome is not stable. | |
Critical appraisal
| Domain | Judgment |
|---|---|
| Residual confounding | Serious risk |
| FindingBefore matching, the propensity-score distributions of the two cohorts diverged clearly, as shown in the supplementary figure, and baseline characteristics separated them on many points: suicidal ideation present in 26.7% of patients directed toward stimulation versus 13.6% under esketamine, a self-harm history in 13.3% versus 5.2%, but nicotine dependence in 13.2% versus 29.0%, diabetes in 11.5% versus 23.2%, prior hospitalization in 37.7% versus 52.8%. The two populations receiving these treatments are not the same. Matching overlays them on the retained variables, with residual standardized differences below 0.10 for all reported characteristics except psychostimulants, at 0.123. It does not restore depressive severity, which was not measured, as the authors acknowledge. | |
| Outcome measurement bias | Serious risk |
| FindingThe monitoring differential between the two treatments produces a differential in opportunities to code an event. The authors’ conclusion, an overall more favorable tolerability profile for stimulation over the first year, cannot be separated from this potential artifact with the data reported, and the authors themselves state this reservation. The negative control softens the objection without removing it, since cellulitis does not have the same probability of being detected during post-administration monitoring as a rhythm disorder or a rise in blood pressure. | |
| Participant selection | Moderate risk |
| FindingThe 1:1 matching yields 1,690 patients per arm, out of 1,802 patients under stimulation and 7,465 under esketamine. It therefore retains 93.8% of the stimulation group and 22.6% of the esketamine group, a calculation made by the editorial team from the published figures. The analyzed esketamine cohort is thus a minority subset, selected to resemble patients treated with stimulation, which shifts the reference population and limits the external validity of the result to patients within the overlap zone of the two propensity scores. The authors also note that the data come primarily from US healthcare organizations. | |
| Multiple analyses | Moderate risk |
| FindingSeveral outcomes, four follow-up windows, two subgroup dimensions, two sensitivity analyses, with no correction mentioned. Several significant results have an interval bound very close to unity: 0.943 for arrhythmia over the whole follow-up, 0.985 for arrhythmia between 14 and 30 days, 1.027 for the suicidal signal among those aged 45 to 65, and 1.031 for that same signal between 30 and 90 days. | |
| Match between claim and evidence | Gap |
| FindingA comparison of efficacy is announced, but efficacy is reduced to coded suicidal outcomes and all-cause care-seeking, with no symptom measure. The discussion passage interpreting lower hospitalization under stimulation as a possible reflection of better overall control of depressive symptoms goes beyond what the outcome measures. To their credit, the authors devote a detailed limitations section, placing unmeasured severity, the indirect nature of the outcomes and the monitoring differential within it, and conclude that a direct randomized trial remains necessary. | |
Level of evidence
Oxford level of evidence 2b, observational cohort study with target trial emulation. Confidence is reasonable on a single point: between day 14 and the end of the first year, no difference in suicidal outcomes was shown between the two treatments, which does not establish their equivalence. It is weak on the somatic tolerability advantage attributed to stimulation, an indeterminate part of which may stem from the difference in monitoring intensity, a reservation the authors state themselves. It is null on symptomatic antidepressant efficacy, which was not measured. The methodological lesson, though, extends beyond the topic: what a target trial emulation corrects, pre-specification, time-zero synchronization, management of immortal time, and what it leaves untouched, confounding by indication on an uncollected variable, is laid bare here.
The colleague test
What an experienced colleague would say if shown this study in two minutes, between two consultations.
You don’t conclude better tolerability by comparing a supervised treatment to one that isn’t, when every outcome is a code from a medical encounter. The authors say so themselves in their limitations. And the word efficacy here doesn’t point to any scale.
Translation for practice: the choice remains open. What can be told to a patient is that no difference was shown on suicidal outcomes over the first year, and that the rest has as much to do with how care is organized as with the treatments themselves.
What you can do with this
- Shared decision-making: both options remain defensible, and this publication does not allow ranking their symptomatic antidepressant efficacy.
- State what is established, namely that no difference appears on suicidal outcomes over the first year, and do not convert this absence of difference into equivalence.
- Bring the practical constraints into the discussion: supervised administration on one side, the frequency and repetition of sessions on the other.
- In a patient over 45 whose suicide risk dominates the clinical picture, keep in mind the subgroup signal favoring esketamine, without turning it into a rule: it is an exploratory, uncorrected analysis whose interval lower bound sits at 1.027 and that the authors themselves ask to be interpreted with caution.
- Safety reminder, outside the scope of the trial: intranasal esketamine is contraindicated in case of vascular aneurysm, arteriovenous malformation or a history of intracerebral hemorrhage, and its administration requires medical supervision; repetitive transcranial magnetic stimulation is contraindicated in the presence of ferromagnetic material or an implanted electronic device near the coil, and calls for caution in case of a seizure history or treatment lowering the seizure threshold. Check these points against the current summary of product characteristics and device instructions for use.
- Keep in mind the detection-bias mechanism, which transfers to any comparison between a supervised treatment and one that is not.
- Wait for a direct randomized comparison before changing practice on the basis of this work, which the authors also request.
Frequently asked questions
Is magnetic stimulation safer than esketamine?
This work does not allow that conclusion. The observed gap on somatic events is confounded with a gap in monitoring intensity, a caveat the authors state in their limitations, and it narrows over time for injuries and arrhythmia. For hospitalizations and emergency visits, however, the gap persists across the whole first year.
What is a target trial emulation actually worth?
It is a rigorous framework that requires writing the protocol before analyzing, fixing time zero and synchronizing eligibility, assignment and the start of follow-up, which removes several classic errors of observational analysis. It does not replace randomization and only balances measured variables, which is precisely the limitation here, since depressive severity was not collected.
Should the suicidal signal between 30 and 90 days be taken seriously?
With caution, which is also the authors’ own position. It is the only significant window on this outcome, within a set of uncorrected multiple comparisons, with an interval lower bound of 1.031. The sensitivity analysis restricted to 2022 through 2025 moves the overall estimate for this outcome from 1.199 to 0.901.
Can a first-line preference be drawn from this?
No. The question remains open, and the scarcity of direct comparisons between these two treatments is the real takeaway of this publication.
Annotated bibliography
Source study. Chen JP, Hsu CW, Chen YT, Fang YY, Tseng PT, Chen MH, Kao YC, Hsu TW, Liang CS. Comparison of rTMS and esketamine for treatment-resistant depression: A target trial emulation. Journal of Affective Disorders, 2026, volume 399, article 121132, issue of April 15, 2026, published online January 6, 2026. DOI 10.1016/j.jad.2025.121132. PMID 41506390. The first two authors contributed equally.
What was consulted. References verified on August 13, 2026 against the full text of the publication and its supplementary material, which includes the STROBE checklist, the queries defining both cohorts, and the figure of propensity-score distributions before and after matching. The figures cited come from the text, tables and figures of the publication; the two retention percentages after matching are a calculation made by the editorial team, flagged as such in the text. The full list of the 50 matched characteristics is not published, the baseline characteristics table presenting only the main ones. Article submitted to an independent double reading.
Editorial collections
Tags
Verified on August 13, 2026 against the full text of the publication and its supplementary material where available. This analysis underwent an independent double reading. The English version was checked for conformity on September 20, 2026, against the figures of the French version and against the source. How we verify what we publish
This analysis is intended for healthcare professionals. It does not constitute a prescribing recommendation and does not replace individual clinical judgment.
Analysis from Psychiatry Evidence Base, evidence-based psychiatry, explained with rigor.