Search strategy and selection criteria

We performed a systematic review and NMA including randomised controlled trials (RCTs) comparing antidepressants with placebo and between each other. Only studies with at least 4 weeks of follow-up were included, as this is a reasonable timeframe within which we expect to observe a clinical response to antidepressants. We searched for RCTs that included adults (≥18 years of age) with a primary diagnosis of one or more physical conditions and a diagnosis of acute depressive episode (including single episode of depression, recurrent depressive disorder, mixed depressive and anxiety disorder or adjustment disorder with clinically relevant depressive symptoms) as assessed by clinical examination with or without the support of validated rating scales (e.g. Hamilton Depression Rating Scale (HDRS), Montgomery–Åsberg Depression Rating Scale (MADRS), Beck Depression Inventory (BDI)) or manualised diagnostic criteria (e.g. the Diagnostic and Statistical Manual of Mental Disorders (DSM) ¹⁹ and the International Classification of Diseases (ICD) ²⁰ ). If the presence of clinically relevant depression was not clearly described, we assessed its presence according to baseline mean scores on validated rating scales measuring depression, considering the following validated cut-offs indicating at least ‘moderate’ depression: HDRS ^{Reference Hamilton21} ≥ 14; MADRS ^{Reference Montgomery and Asberg22} ≥ 20; BDI version I (BDI-I) ^{Reference Beck, Ward, Mendelson, Mock and Erbauch23} ≥ 10 and version II (BDI-II) ^{Reference Beck, Steer and Brown24} ≥ 14; Major Depression Inventory (MDI) ^{Reference Bech, Timmerby, Martiny, Lunde and Soendergaard25} ≥ 26; Center for Epidemiologic Studies Depression Scale (CES-D) ^{Reference Lewinsohn, Seeley, Roberts and Allen26} ≥ 16; Brief Zung Self-Rating Depression Scale (BZSRD) ^{Reference Zung27} ≥ 44; Hospital Anxiety and Depression Scale (HADS) ^{Reference Snaith and Zigmond28} ≥ 11; Patient Health Questionnaire (PHQ-9) ^{Reference Kroenke, Spitzer and Williams29} ≥ 10; Symptom Checklist-90-Revised (SCL-90-R) ^{Reference Derogatis, Lipman and Covi30} depression subscale ≥ 20. In the case of studies with mixed populations (i.e. depression and anxiety), we included only those with at least 80% of participants suffering from clinically relevant depression, as defined above.

We excluded conditions for which a clear pathogenesis could not be identified (e.g. physically unexplained symptoms, functional symptoms), or conditions for which a relevant psychiatric or psychological component may be particularly relevant (e.g. somatic symptoms disorders, fibromyalgia, irritable bowel syndrome). We also excluded minor depression, dysthymia, prolonged grief disorder or complicated bereavement. No exclusion criteria were applied in terms of setting (e.g. in- and out-patient settings; physical, surgical, psychiatric setting) or type of antidepressant, as even those that are uncommonly prescribed or no longer marketed (e.g. nomifensine) can indirectly contribute to estimate the differential effect between other antidepressants.

We searched electronic databases (MEDLINE, Embase, PsycINFO, CENTRAL, CINAHL), databases of regulatory agencies (e.g. Food and Drug Administration, European Medicines Agency) and trial registers (e.g. clinicaltrials.gov, the World Health Organization’s International Clinical Trials Registry Platform) from inception until 30 April 2024, without language restrictions (see Supplement B for the full research syntax). Two review authors (B.D.L., A.C.) independently assessed titles and abstracts of all retrieved records, and then the full texts of all potentially eligible records. Disagreements were resolved by discussion and consensus with a third review author (G.O.). Data extraction was performed between 1 June 2024 and 15 July 2024 by two review authors (B.D.L., A.C.) independently, in agreement with the recommendations of the Cochrane Handbook for Systematic Reviews of interventions. ^{Reference Higgins, Thomas, Chandler, Cumpston, Li and Page31} Two review authors (G.V., A.C.) independently assessed the risk of bias of included studies using the Cochrane Risk-of-Bias tool, version 2 (RoB2). ^{Reference Sterne, Savović, Page, Elbers, Blencowe and Boutron32} Disagreements were resolved by discussion and consensus with a third review author (G.O., C.B.).

Outcomes

The primary outcomes were (a) depressive symptoms measured as the mean change score (or, alternatively, the end-point score) ^{Reference Ostinelli, Efthimiou, Luo, Miguel, Karyotaki and Cuijpers33} on validated rating scales (HDRS, BDI, MADRS or any other) by the end of the trial, and (b) tolerability, defined as the number of individuals withdrawing from treatment because of adverse events by the end of the trial, as a proportion of the total number of randomised participants.

Secondary outcomes included the following dichotomous outcomes: response; remission; acceptability; death caused by worsening of the physical condition; death from any cause; serious adverse events (SAEs); and the following continuous outcomes: anxiety symptoms; quality of life; social functioning.

Data analysis

Analyses were conducted with Stata (version 18.5 for macOS), and with the R programming language (R version 4.4.0 for macOS; R Foundation for Statistical Computing, Vienna, Austria; https://cran.r-project.org/) within the R Studio integrated development environment. For each outcome and comparison, we performed a standard pairwise, random-effects meta-analysis and a NMA with a random-effects model in a frequentist framework, using the R packages netmeta and the Stata package mvmeta. ^{Reference Balduzzi, Rücker, Nikolakopoulou, Papakonstantinou, Salanti, Efthimiou and et al34,Reference White35} For dichotomous outcomes, risk ratios were log-transformed and pooled using a conventional normal-normal random-effects model with 95% confidence intervals, applying a strict intention-to-treat (ITT) approach (all randomly assigned participants as the denominator). For continuous outcomes, we pooled mean differences if all trials used the same rating scale; otherwise, we employed standardised mean differences (SMDs), including all participants that contributed to the analysis in the original trial. We calculated SMDs as the ratio between the mean difference between groups and the standard deviation of outcome among participants, according to the Cochrane Handbook. ^{Reference Higgins, Thomas, Chandler, Cumpston, Li and Page31} If we could not retrieve missing data from trial authors, we imputed them with validated statistical methods, including imputing response and remission according to mean rating scale scores, standard deviations and validated cut-offs. ^{Reference Higgins, Thomas, Chandler, Cumpston, Li and Page31,Reference Furukawa, Barbui, Cipriani, Brambilla and Watanabe36,Reference Furukawa, Cipriani, Barbui, Brambilla and Watanabe37}

For the NMA, common heterogeneity across all comparisons was assumed in each network, and global heterogeneity was assessed using τ² (low, τ ² ≤ 0.010; moderate, 0.010 < τ ² ≤ 0.242; and high, τ ² > 0.242), estimated by the DerSimonian–Laird method through the R netmeta function ^{Reference Huhn, Nikolakopoulou, Schneider-Thoma, Krause, Samara and Peter38} and I² (low, 0–40%; moderate, 30–60%; substantial, 50–90%; and considerable, 75–100%). ^{Reference Higgins, Thomas, Chandler, Cumpston, Li and Page31}

To assess the assumption of transitivity, we compared the distribution of the following variables across treatment strategies: sample size; year of publication; follow-up duration; blinding (double blind versus open label); setting (in-patient, out-patient or mixed settings); industry sponsorship; mean age; percentage of female participants; mean psychopathology score at baseline; mean duration of depression; depression as the primary study outcome; mean number of co-occurring physical conditions; mean severity of the physical condition at baseline (according to a classification derived from the Severity of Illness Instrument). ^{Reference Horn, Horn and Sharkey39}

For the primary outcomes, we visually inspected boxplots on the distribution of continuous variables across treatments and assessed whether imbalances were large enough to threaten the transitivity assumption according to both the Kruskal–Wallis test and meta-regression analyses.

We evaluated the presence of inconsistency, defined as the statistical disagreement between direct and indirect evidence of a treatment comparison, by comparing direct and indirect evidence within each closed loop and comparing the goodness of fit for a NMA model that assumes consistency with a model that allows for inconsistency in a ‘design-by-treatment interaction model’ framework by using the R Studio netmeta package and decomp.design and netsplit commands. ^{Reference Balduzzi, Rücker, Nikolakopoulou, Papakonstantinou, Salanti, Efthimiou and et al34,Reference Horn, Horn and Sharkey39} Where we found evidence of inconsistency, we investigated this further using both a node-splitting and a side-splitting approach between comparisons. For each outcome, we calculated the probability of each antidepressant of being at each possible rank and the treatment hierarchy by means of p-scores (mean normalised rank probabilities), which reflect the extent of certainty that a treatment is better than competing treatments, based on the point estimates and standard errors. ^{Reference Rücker and Schwarzer40} If a comparison included ten studies or more, we assessed publication bias by visually inspecting the funnel plot, tested them for asymmetry with the Egger’s regression test and investigated possible reasons for funnel plot asymmetry.

We performed sensitivity analyses excluding RCTs with the following characteristics: placebo controlled; not double blind; with overall high risk of bias according to RoB2; with follow-up shorter than 3 months; efficacy on depressive symptoms was not the primary study aim; without a formal diagnosis of major depression (post hoc); with a small sample size (<50 participants; post hoc analysis); recruiting participants with illness of less than moderate severity. Furthermore, we performed meta-regression analyses on the same variables assessed for transitivity, to identify possible treatment effect moderators.

We also performed the following subgroup analyses for the primary outcomes, provided that at least three RCTs can be pooled together: analysing each subgroup of diseases (according to the ICD-11 classification) and each physical condition separately; grouping treatments according to their class, namely selective serotonin reuptake inhibitors (SSRIs), selective serotonin and noradrenaline reuptake inhibitors (SNRIs), tricyclic antidepressants (TCAs) and other antidepressants.

For the primary outcomes, certainty of the pooled evidence was assessed using the Confidence in Network Meta-Analysis (CINeMA) approach. ^{Reference Nikolakopoulou, Higgins, Papakonstantinou, Chaimani, Del Giovane and Egger41}