Introduction
Despite major depressive disorder (MDD) being highly prevalent [1–Reference Epskamp, Borsboom and Fried3], the neurobiological underpinnings remain under investigation and treatment options are still suboptimal [Reference Moncrieff, Cooper, Stockmann, Amendola, Hengartner and Horowitz4, Reference Kim, Kim, Lee and Koo5]. Effective treatment options include antidepressants (AD) targeting monoamine reuptake and receptor binding [Reference Cipriani, Furukawa, Salanti, Chaimani, Atkinson and Ogawa6], as well as glutamatergic signaling and neuroplasticity [Reference Krystal, Kavalali and Monteggia7]. Nevertheless, many patients do not show sufficient symptom relief even after multiple treatment trials and develop treatment-resistant depression (TRD) [Reference McIntyre, Alsuwaidan, Baune, Berk, Demyttenaere and Goldberg8]. TRD has been extensively studied and linked to various genetic [Reference Fabbri, Kasper, Kautzky, Bartova, Dold and Zohar9, Reference Fabbri, Corponi, Souery, Kasper, Montgomery and Zohar10], environmental [Reference Fabbri, Hosak, Mossner, Giegling, Mandelli and Bellivier11], and clinical characteristics [Reference Kautzky, Dold, Bartova, Spies, Kranz and Souery12]. The framework of the research consortium “European group for the study of resistant depression” (GSRD), consisting of consecutive waves of patients with severe MDD whose treatment was outcome assessed across multiple trials, has contributed risk factors for TRD [Reference Bartova, Dold, Kautzky, Fabbri, Spies and Serretti13, Reference Kautzky, Moller, Dold, Bartova, Seemuller and Laux14]. However, there are still no validated clinical tools for early risk assessment of TRD and, more importantly, optimal treatment choices to prevent TRD. Heterogeneity in the neurobiological and environmental pathways leading to depression have been identified as obstacles in finding consistent diagnostic and prognostic markers in depression. A related important factor is the lack of a representation of the plethora of depressive symptoms and their respective contributions to the variability in treatment outcomes [Reference Fried15].
Targeting depression on the symptom level and finding patterns of co-expression has gained interest in recent years [Reference Fried16]. Symptom networks highlight influential symptoms that are central to the clinical presentation of MDD [Reference Borsboom17], distinct to a subgroup of patients, or specific in response to treatment [Reference Epskamp2]. Results were shown to replicate and may bring more precision to treatment choices [Reference Borsboom, Deserno, Rhemtulla, Epskamp, Fried and McNally18]. However, studies comparing network characteristics before and after antidepressant treatment are still scarce [Reference Bos, Fried, Hollon, Bringmann, Dimidjian and DeRubeis19–Reference Madhoo and Levine21] and implications of symptom networks remain under discussion [Reference Hoekstra, Epskamp and Borsboom22, Reference Epskamp and Isvoranu23]. The concept of TRD itself was recently challenged as empty [Reference Braillon, Fried, Cristea, Cosgrove and Naudet24], however, symptom patterns specific to TRD that may inform targeted treatment have not been analyzed yet. Thus, we aimed to compare symptom networks in treatment response, non-response, and TRD in patients with MDD and to identify symptoms central to specific outcomes.
Methods
Sample
The TRD-III cohort of the GSRD project was used for network generation [Reference Bartova, Dold, Kautzky, Fabbri, Spies and Serretti13], including a total of 1,410 patients with MDD recruited 2011–2016 as in- or outpatients in 15 community and university hospitals across Europe and Israel. Diagnoses were established with the Mini International Neuropsychiatric Interview (MINI) and MDD was required to be the primary diagnosis [Reference Sheehan, Lecrubier, Sheehan, Amorim, Janavs and Weiller25]. Depression severity was assessed using the Montgomery-Åsberg-depression rating scale (MADRS) for two timepoints [Reference Montgomery and Asberg26]: pretreatment and when response, non-response, or TRD were established. The pretreatment MADRS score was assessed retrospectively based on medical records and patients´ recollection of symptom severity at the time of treatment initiation. At least moderate depression severity according to a pretreatment MADRS score > 21 was required.
The European staging system for treatment outcome was used [Reference Souery, Amsterdam, de Montigny, Lecrubier, Montgomery and Lipp27]:
Treatment response was defined by a decline ≥50% in MADRS score after an adequate antidepressant treatment trial compared to pretreatment scores assessed retrospectively. If response was not reached after a single AD trial of adequate length and dosage, outcome was labeled non-response. If response was not achieved after two or more antidepressant trials were administered, outcome was labeled TRD. Detailed descriptions of the TRD-III sample regarding clinical characteristics and outcomes have been published previously [Reference Bartova, Dold, Kautzky, Fabbri, Spies and Serretti13].
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2013. All procedures involving human subjects/patients were approved by the Ethics Committee of the Medical University of Vienna (EK 1133/11).
Network estimation
Symptom networks were estimated with the “bootnet” package for the statistical program “R” (version 4.3.0) [Reference Epskamp, Borsboom and Fried3]. Following published recommendations [Reference Epskamp and Fried28], matrices of Spearman correlations between each symptom pair were used to estimate sparse Gaussian graphical models. Nodes correspond to depressive symptoms according to the 10 MADRS items and edges to their partial correlation. The mean score of items 1 and 2, both addressing sadness, was used, totaling 9 nodes and a maximum of 36 interconnecting edges. The most parsimonious model, i.e., the model with optimal performance while minimizing the number of edges, was determined by optimization of the extended Bayesian information criterium by graphic L1 regularization (EBIC-lasso). The default parameter of λ = 0.5 was applied. Network stability and accuracy of edge weights were assessed by bootstrapping following published recommendations for “bootnet” [Reference Epskamp, Borsboom and Fried3]. Detailed information on bootstrapping procedures and assessment of Berkson’s bias can be found in Supplementary Section [Reference de Ron, Fried and Epskamp29].
Next, strength was assessed for each node of a network by summing up the weights of all connected edges. Thus, node strength indicates the centrality of a specific symptom regarding all other symptoms of a network.
Networks were then visualized with the “qgraph” package [Reference Epskamp, Cramer, Waldorp, Schmittmann and Borsboom30]. Thereby, all edges estimated by EBIC-lasso were visualized. Edge weights, ranging between 0 and 1, were labeled for each stable edge according to bootstrapping. In addition, strength indices for each node of each network were visualized as line plots.
Network comparison
To assess differences in network structures between each pair of networks, the R-package “NetworkComparisonTest” (NCT) was used to compute the statistics S (difference in global network strength) and M (network invariance assessed by maximum difference in edge weights) [Reference Van Borkulo, Van Bork, Boschloo, Kossakowski, Tio and Schoevers31]. We compared between-subjects networks estimated for each treatment outcome group, respectively pre- and posttreatment, as well as within-subjects networks for symptoms reported pre- and posttreatment within each outcome group. The latter comparisons were computed with the experimental “paired” option of NCT. Furthermore, for each significant comparison of two networks according to S and M statistics, differences in individual edges (whenever M < 0.05) and node strength (whenever S < 0.05) were tested. Edges were compared only when they were stable according to bootstrapping in both networks. For each NCT result, uncorrected p-values are reported with a significance level < 0.05.
Results
MADRS ratings were fully available for 1,385 patients (66.8% female, mean age 51.7 ± 14.1 years), including 326 responders (pretreatment MADRS 32.6 ± 8.0), 489 non-responders (pretreatment MADRS 33.2 ± 8.1), and 570 patients with TRD (pretreatment MADRS 35.8 ± 6.7). Details on age, sex, and presentation of depressive symptoms both pre- and posttreatment are provided in Table 1. The distribution of MADRS sum scores and individual items, pre- and posttreatment, is displayed in Supplementary Figures 1–3. For a distribution of severe symptoms (MADRS item score > 3), for each outcome and pre- and posttreatment, see Supplementary Table 1.
Sex, age, and symptom presentation for the three treatment outcomes, shown both pre- and posttreatment
Note: Means are shown with standard deviations; medians, with range.
Network structure pretreatment
Stability of edge weights was good for networks estimated, respectively for patients with TRD and non-responders, and acceptable in the group of treatment responders (Figure 1). Stability of node strength was good in non-responders and acceptable in patients with TRD and responders. Networks estimated on symptoms pretreatment are depicted in Figure 2A. No significant differences were detected by NCT regarding global strength, ranging from 3.5 in responders to 3.9 in TRD patients (all p > 0.05). Network invariance was not significantly different between networks in TRD and non-responders, but it was different between responders and TRD (M = 0.21, p = 0.043), and between responders and non-responders (M = 0.21, p = 0.037). No significant differences were observed between individual stable edges.
Case-dropping subset bootstrap results. On the x-axis the proportion of dropped cases is shown, while the y-axis indicates the lower 95% confidence interval (CI) of the correlation coefficient between the 1,000 bootstrapped samples (termed correlation stability (CS) coefficient), respectively for (A) node strengths and (B) edge weights. The cutoff of 0.7 is marked by a dotted line and colored areas under this cutoff indicate poor, acceptable, and good stability. Colored lines indicate the achieved correlation coefficients for each network, for pre- and posttreatment and the phenotypes of treatment-resistant depression (TRD), non-response, and response.
Gaussian graphical models estimated for patients with TRD, non-, and response to treatment for two timepoints, (A) before initiation of antidepressant treatment, and (B) after treatment outcome was determined. Positive partial correlations, i.e., co-expression of severity of two symptoms, are portrayed in blue color while negative partial correlations, i.e., high load of one symptom occurring with low severity of another, are portrayed in red color. In panel (C), standard deviations from mean edge weights of the three networks estimated for pre- and posttreatment are displayed. Colors correspond to presence of edges in specific or multiple outcome phenotypes. In all networks, edge weights are only displayed when stable according to bootstrapping. Nodes are colored dark to light by declining node strength, i.e., the sum of edge weights connecting each node. Abbreviations: APP, reduced-appetite; CON, concentration difficulties; ItF, inability-to-feel; LAS, lassitude; PES, pessimistic thoughts; SAD, sadness; SLE, reduced-sleep; SUI, suicidal thoughts; TEN, tension; TRD, treatment-resistant depression.
Stable edges between lassitude sadness, concentration difficulties and inability-to-feel, as well as between pessimistic thoughts and suicidal thoughts, were shared between all baseline networks (Figure 2B). Within all three networks, the strongest edges as well as nodes were not significantly different from each other, but only compared to the lower-ranking edges and nodes (Supplementary Figure 4).
In patients with TRD, a network with 32 edges (18 stable) was estimated. The strongest edge was observed between tension and concentration difficulties, which was unique to TRD. Regarding node strength, pessimistic thoughts ranked before sadness, lassitude, inability-to-feel, and concentration difficulties (Figure 3A). When considering only stable edges (Figure 3B), lower node strength was observed for pessimism.
Symptom centrality assessed by node strength, that is, the sum of edge weights connecting a specific symptom directly to other symptoms. Results are colored by treatment outcomes TRD, non-response, and response. Panel (A) shows strength rankings for edges, while panel (B) shows rankings when restricting strength to stable edges with non-zero confidence intervals according to bootstrapping. APP, reduced-appetite; CON, concentration difficulties; ItF, inability-to-feel; LAS, lassitude; PES, pessimistic thoughts; SAD, sadness; SLE, reduced-sleep; SUI, suicidal thoughts; TEN, tension; TRD, treatment-resistant depression.
In patients with treatment non-response 30 edges (17 stable) were estimated, with the strongest edge between concentration difficulties and inability-to-feel, which was significant also in the groups of TRD and responders. Unique edges linked tension to sleep, pessimistic thoughts, and suicidal thoughts. Regarding node strength, sadness ranked before lassitude, inability-to-feel, suicidal thoughts, and pessimistic thoughts. When considering only stable edges, concentration difficulties ranked before pessimistic thoughts.
In patients with treatment response, 30 edges (13 stable) were estimated. The strongest edge linked reduced-appetite to tension, which was unique to the response network. A second unique edge was found between reduced-sleep and lassitude. Highest node strength was found for reduced-appetite, lassitude, suicidal thoughts, concentration difficulties, and inability-to-feel. A lower ranking was observed for reduced-appetite when considering only stable edges.
Network structure posttreatment
Stability of both edge weights and node strength was good for networks estimated for patients with TRD and non-responders, and acceptable in the group of treatment responders (Figure 1). Networks are depicted in Figure 2C. Networks differed in global strength, ranging from 2.3 in responders to 3.2 in non-responders (all p < 0.05). For individual node strengths differing between networks, please see Table 2, panel B. Structure was not different between any networks (all p > 0.05).
Results of the network comparison tests, respectively for cross-sectional comparisons between treatment outcomes and longitudinal comparisons within each treatment outcome
Note: (A) Global tests included network invariance, that is, whether network structures are different, as well as network strength, that is, whether node strengths are different. (B) In case of significant global results, individual edges or nodes were tested.
n.s. not significant
Generally, fewer edges unique to any single network were observed, particularly when considering edges between high-centrality symptoms of sadness, lassitude, and inability-to-feel, which were stable across all networks. Edges connecting to pessimistic thoughts were shared in TRD and non-responders. In all three networks, the strongest edges as well as nodes were not significantly stronger than each other, but only compared to the low-ranking edges and nodes (Supplementary Figure 5).
Networks with 27 edges (14 stable) and 20 edges (10 stable) were estimated in TRD and non-response. The strongest edges in TRD linked pessimistic thoughts to sadness and lassitude, while the strongest edge in non-response linked concentration difficulties to inability-to-feel. A unique edge was observed in non-responders, linking reduced-appetite to suicidal thoughts. In both TRD and non-responders, sadness ranked before lassitude, inability-to-feel, and pessimistic thoughts, followed by concentration difficulties in non-responders and tension in TRD. When considering only stable edges, relatively higher node strength was observed for lassitude and pessimism in TRD.
In patients with response, 20 edges (10 stable) were estimated. The strongest edge was estimated between sadness and pessimistic thoughts and a unique edge was observed between reduced-appetite and inability-to-feel. The highest node strength was observed in sadness, inability-to-feel, lassitude, concentration difficulties, and pessimism, irrespective of considering all or only stable edges.
Longitudinal changes in network structure
Across treatment, both in responders and in TRD network strength declined (TRD: S = 1.17, p < 0.001; response: S = 1.16, p = <0.001) and structure changed (TRD: M = 0.2, p < 0.001; response: M = 0.26, p = 0.026). In contrast, in non-responders, no change in global strength or invariance was observed. Edges unique to specific phenotypes pretreatment either lost significance or were shared between networks posttreatment (Figure 2B).
Changes in edge weights for each treatment outcome are displayed in Supplementary Figure 6. In TRD, biggest reductions and loss of stability were observed for edges linking reduced-appetite and reduced-sleep, concentration difficulties and inability-to-feel, and pessimistic thoughts and suicidal thoughts. In contrast, the link from tension to pessimistic thoughts increased and reached stability posttreatment. Among edges that were stable both pre- and posttreatment, the link from tension to concentration difficulties was reduced significantly (Δ = 0.14, p = 0.01), while small increases were observed between pessimistic thoughts and sadness (Δ = 0.02, p > 0.05) and lassitude (Δ = 0.04, p > 0.05).
In responders, biggest reductions in edge weights were observed between reduced appetite and tension, between reduced appetite and concentration difficulties, between reduced appetite and suicidal-thoughts, between suicidal-thoughts and pessimistic-thoughts, between suicidal-thoughts and sadness, between lassitude and reduced-sleep, and between lassitude and pessimistic-thoughts. Edges increased between sadness and pessimistic-thoughts as well as between tension and both sadness and concentration difficulties, reaching stability only post-treatment.
Discussion
We identified shared and outcome-specific changes in symptom connectivity when applying psychometric network modeling to a naturalistic cohort of 1,385 patients with MDD and treatment-response, non-response, and TRD.
Global network strength was not significantly different between groups pretreatment, albeit the highest scores were observed in TRD followed by non-response and lastly response. On one hand, higher symptom connectivity was associated with unfavorable outcomes such as higher risk for MDD onset [Reference Cramer, van Borkulo, Giltay, van der Maas, Kendler and Scheffer32], symptom persistence [Reference van Borkulo, Boschloo, Borsboom, Penninx, Waldorp and Schoevers20], treatment non-response [Reference McElroy, Napoleone, Wolpert and Patalay33, Reference Lee, Kelley, Palacios, Richards and Gillan34], or relapse following successful psychotherapy [Reference Lorimer, Delgadillo, Kellett and Brown35]. Contrary to prior suggestions of differences between outcome groups being driven by higher variance of mean symptom scores rather than phenotypic qualities [Reference Lee, Kelley, Palacios, Richards and Gillan34], here lowest variance of average symptom load but highest global strength was observed in TRD. However, results must be interpreted cautiously as we cannot rule out that retrospective assessment of pretreatment symptoms introduced a bias to severity ratings.
On the other hand, increasing network strength throughout treatment was reported [Reference McElroy, Napoleone, Wolpert and Patalay33, Reference Berlim, Richard-Devantoy, Dos Santos and Turecki36, Reference Beard, Millner, Forgeard, Fried, Hsu and Treadway37], likely due to symptoms decreasing simultaneously following successful treatment. However, some studies suggested persistence of network strength [Reference Kaiser, Herzog, Voderholzer and Brakemeier38], depending on specific symptom associations [Reference Schumacher, Klein, Elsaesser, Harter, Hautzinger and Schramm39]. Here, global strength was reduced across all groups following at least 4 weeks of antidepressant treatment. Thereby, posttreatment networks became more similar across groups, indicated by fewer edges being unique to a specific outcome and converging rankings regarding symptom strength. Interestingly, significant reductions in strength were found both in TRD and response but not in non-response. From a clinical perspective, the non-response network resembles an earlier stage of treatment compared to TRD patients who were treated for ≥4 additional weeks. Our results are thus suggestive of decreased global network strength relative to severity reduction and treatment duration. Consequently, some aspects of the TRD phenotype may be produced by partial response to sequential antidepressant trials that cure more easily treatable symptoms while those more resistant remain. Particularly in the case of unfavorable outcomes, connections between sadness, lassitude, inability-to-feel, and pessimistic thoughts densified posttreatment. While dense connections could theoretically also indicate a shared decrease of these four symptoms, here we observed high rates of persistence. Lassitude was frequent and severe specifically in the cohort of TRD patients, both pretreatment (83% with MADRS item score > 3, compared to 64–67% in other phenotypes) and posttreatment (90.7% with MADRS item score ≥ 3). Similarly, across all outcome phenotypes but particularly in TRD, high symptom loads of sadness were observed pretreatment (94% with MADRS item >3 in TRD, 78–82% in other phenotypes) and sadness showed higher rates of persistence than any other symptom (96.3% with MADRS item score ≥ 3 in TRD). The increased connectivity between these core symptoms despite the reduction of overall symptom load therefore seems to indicate resistance, and it may have higher clinical relevance than global changes in network strength.
Previous research on the factor structure of MADRS suggested one to four factors, but more recent analyses in large samples of severely depressed patients with MDD agreed on a four-factor model, including factors sadness, neurovegetative symptoms, detachment, and negative thoughts [Reference Quilty, Robinson, Rolland, Fruyt, Rouillon and Bagby40, Reference Seemuller, Schennach, Musil, Obermeier, Adli and Bauer41]. This model presented with configural and metric invariance over antidepressant treatment, indicating that items maintained the direction and magnitude of their factor loadings. However, these results were based on samples with higher rates of treatment response and TRD has thus far not been investigated in this context. Reflecting on the current results, longitudinal factor patterns may differ between treatment outcome groups as patients with TRD appear to maintain higher symptoms related to factors “sadness” (sadness), “detachment” (lassitude and inability-to-feel), and “negative thoughts” (pessimistic thoughts) compared to other symptoms. These symptoms partly align with the proposed anhedonia factor of the MADRS, which also includes concentration difficulties while omitting pessimistic thoughts [Reference McIntyre, Necking, Schmidt and Reines42].
Regarding individual symptom scores, sadness, lassitude, pessimistic thoughts, and inability-to-feel were the most densely connected symptoms. While outcome-specific rankings of node strength were observed pretreatment, posttreatment these four symptoms showed the highest connectivity across all outcomes. Our results thus agree with a recent review on MDD symptom networks that identified fatigue as the top-ranked symptom regarding centrality strength [Reference Malgaroli, Calderon and Bonanno43]. Fatigue is a symptom describing reduced energy and exhaustion, thus corresponding well with lassitude assessed by the MADRS as in our study. Fatigue is also a symptom known for high persistence throughout treatment and as a residual symptom associated with the risk of relapse [Reference Fava, Ball, Nelson, Sparks, Konechnik and Classi44]. Strong connections linked lassitude to sadness and concentration difficulties across all outcomes, while a strong connection to pessimistic thoughts was found exclusively in TRD and particularly posttreatment. Fatigue was previously suggested to be part of somatic symptoms of depression, bridging them to emotional symptoms via anhedonia [Reference Malgaroli, Calderon and Bonanno43]. In line with this hypothesis, strong links were identified with inability-to-feel, which can be regarded as a representation of anhedonia, but also directly to sadness as well as to concentration difficulties. Anhedonia symptoms placing centrally all outcomes may also suggest that this subgroup of patients could benefit from combination treatments targeting anhedonia to possibly reduce both TRD risk and residual symptoms in responders [Reference Serretti45].
Sadness, the other symptom consistently leading centrality rankings in MDD [Reference Kaiser, Herzog, Voderholzer and Brakemeier38, Reference Malgaroli, Calderon and Bonanno43], was the most strongly connected symptom posttreatment, agreeing with previous findings in the STAR*D sample [Reference Madhoo and Levine21]. Surprisingly, sadness placed below decreased-appetite and suicidal thoughts in centrality rankings of the pretreatment symptom network solely in the responder group. Both appetite changes and suicidality are typically placed among the least strongly connected symptoms in network analyses in MDD [Reference Malgaroli, Calderon and Bonanno43], which is also in line with our previous finding of HAM-D symptom reduced-appetite clustering only with reduced-weight but no other emotional or somatic symptoms in two naturalistic cohorts of depressed patients [Reference Kautzky, Moller, Dold, Bartova, Seemuller and Laux14]. However, decreased appetite can be easily targeted by treatment, and response in this symptom domain was more easily predictable by machine learning classification models than other symptoms such as emotional or anxious symptoms [Reference Kautzky, Moller, Dold, Bartova, Seemuller and Laux14, Reference Serretti and Mandelli46]. On a more speculative note, acute suicidality may trigger more targeted treatments such as ketamine, lithium augmentation, or electroconvulsive therapy, which were prescribed to some patients as part of the naturalistic treatment in the present patient cohort [Reference Shamabadi, Ahmadzade and Hasanzadeh47–Reference Zalsman, Hawton, Wasserman, van Heeringen, Arensman and Sarchiapone49]. Consequently, a relatively lower strength of core symptoms such as sadness, lassitude, and inability-to-feel compared to other symptoms that could either be more directly ameliorated by treatment or guide more invasive treatment options may be indicative of higher chances for treatment response. However, following this hypothesis, also reduced-sleep could be expected to show higher centrality in treatment response, as it can be specifically targeted by sedating antidepressants and antipsychotics, but the opposite association was reported by previous research [Reference Lee, Kelley, Palacios, Richards and Gillan34].
In summary, while previous discussions about the indications of centrality measures generally argued in favor of more central symptoms being more relevant to treatment [Reference Fried, Epskamp, Nesse, Tuerlinckx and Borsboom50], mechanisms behind this assumption may be less generalizable and dependent on specific symptom connections. On one hand, a symptom that is more densely connected to other symptoms can be a prime treatment target as lowering the symptom load may also reduce concomitant symptoms. However, more isolated symptoms with high perseverance throughout treatment may be particularly difficult to cure, as synergistic effects from neighboring symptoms are missing. The mechanisms involved may be dependent on the specific symptom edges present in an individual patient. As an example, a link between suicidality and worthlessness, the latter being here represented by pessimistic thoughts, was commonly observed in network analyses targeting MDD and also consistently observed here. However, this link was weaker in TRD compared to other outcome groups and lost significance in the TRD group after treatment. Whether the presence and strength of this connection has direct implications for treatment of either of the two symptoms remains to be resolved.
While we followed published recommendations regarding network generation and testing for stability and accuracy and took into consideration potential pitfalls such as Berkson’s bias, we cannot rule out spurious findings as the generalizability of the network results could not yet be validated in an independent sample. While a methodological framework for applying network structures to novel datasets has been published [Reference Epskamp and Isvoranu23], to our knowledge there are no publicly available samples specifically addressing TRD. Further, limitations specific to the TRD-III data set must be considered. Results are based on a naturalistic sample of highly heterogeneous patients with a broad range of interventions applied in different treatment centers across Europe. While all interviews for MINI and MADRS were performed by specialist medical staff with training, inter-rater reliability was not tested and differences in ratings may have influenced the results. This concern is somewhat relieved by our prior studies based on the TRD-III dataset that showed consistent performance of models based on clinical characteristics when a cross-center validation design was applied [Reference Kautzky, Dold, Bartova, Spies, Vanicek and Souery51]. Most importantly, medical documentation of baseline symptoms was not routinely available and the retrospective assessment of pretreatment MADRS scores relied on adequate recollection of the participating patients. While prior studies have shown good reproducibility regarding retrospective assessment of symptom types [Reference Dunlop, Granros, Lechner, Mletzko-Crowe, Nemeroff and Mayberg52], recollection of symptom severity may be biased by the current disease state. On one hand, patients achieving response may have better insight and find it easier to identify past symptoms such as reduced-appetite. On the other hand, unsuccessful treatment outcomes and persistent symptom load may lead to overestimation of initial severity. However, mean baseline symptom scores were not statistically significant between groups, and also results regarding group differences in baseline networks are in line with clinical considerations and previous research [Reference Malgaroli, Calderon and Bonanno43]. Nevertheless, retrospective symptom assessment certainly introduces qualitative differences in networks estimated for pre- and posttreatment that cannot be accounted for by statistical methods. Particularly these results must therefore be regarded preliminary and need further validation by longitudinal study designs.
In synopsis, we observed differences in network structure between patients with MDD and treatment response, non-response, and TRD for symptoms reported both pretreatment and posttreatment. Symptom connections including pessimistic thoughts were particularly pronounced in TRD, while symptoms of reduced-appetite and suicidal thoughts were more densely connected in the response group. Posttreatment, global strength was higher in patients with TRD and in non-responders compared to responders, but network structures became more similar and significant edges were mostly shared between outcome groups. Despite overall decreases in network strength following treatment, specific connections between core symptoms of sadness, lassitude, inability-to-feel, and pessimistic thoughts increased. Particularly in TRD, these symptoms showed high rates of persistence after treatment and patients may benefit from targeted treatment approaches.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1192/j.eurpsy.2025.2454.
Data availability statement
The data that support the findings of this study contain sensitive information that could compromise the privacy of research participants and are therefore are not publicly available. The code used to generate the results as well as a randomly generated sample similar in design to the original data is available as Supplemental Material. The original data that were used to support the findings of this study are available from the corresponding author S.K. upon reasonable request.
Acknowledgements
The authors would like to thank all persons involved in the GSRD project and the patients who participated in the present study.
Author contribution
Dr. Kautzky and Dr. Kasper were significantly involved in designing the study and implementation of the research. Dr. Kautzky performed the statistical analyses and wrote the report including the first draft of the manuscript that was further elaborated and critically revised by Dr. Kasper. All authors substantially contributed to implementation of the research and have critically revised and approved the final manuscript.
Financial support
The European Group for the Study of Resistant Depression (GSRD) obtained an unrestricted grant sponsored by Lundbeck A/S. The sponsor played no role in designing the study, data collection and analyses, interpretation of the data, writing of the manuscript, and in the decision to submit the research for publication.
Competing interests
Within the last 3 years, Dr. Bartova has received travel grants and consultant/speaker honoraria from Alpine Market Research, Angelini, Biogen, Diagnosia, Dialectica, EQT, Janssen, Lundbeck, Market Access Transformation, Medizin Medien Austria, Novartis, Schwabe, and Universimed. Dr. Dold has received travel grants and consultant/speaker honoraria from Medizin Medien Austria, Janssen, and Universimed. Dr. Fugger has received consultant/speaker honoraria from Janssen and Schwabe. Dr. Rujescu has received grant/research support from Janssen and Lundbeck; he has served as a consultant or on advisory boards for Janssen and Rovi and he has served on speakers bureaus of Janssen and Pharmagenetix. Dr. Souery has received grant/research support from GlaxoSmithKline and Lundbeck; and he has served as a consultant or on advisory boards for AstraZeneca, Bristol-Myers Squibb, Eli Lilly, Janssen, and Lundbeck. Dr. Mendlewicz is a member of the board of the Lundbeck International Neuroscience Foundation and of the advisory board of Servier. Dr. Zohar has received grant/research support from Lundbeck, Servier, and Pfizer; he has served as a consultant or on the advisory boards for Servier, Pfizer, Solvay, and Actelion; and he has served on speakers’ bureaus for Lundbeck, GlaxoSmithKline, Jazz, and Solvay. Dr. Montgomery has served as a consultant or on advisory boards for Lundbeck. Dr. Fabbri has served as speaker for Janssen. Dr. Serretti has served as a consultant or speaker for Abbott, Abbvie, Angelini, AstraZeneca, Clinical Data, Boehringer, Bristol-Myers Squibb, Eli Lilly, GlaxoSmithKline, Innovapharma, Italfarmaco, Janssen, Lundbeck, Naurex, Pfizer, Polifarma, Sanofi, Servier, and Taliaz. Within the last 3 years, Dr. Kasper has received grant/research support from Lundbeck; he has served as a consultant or on advisory boards for Angelini, Biogen, Esai, Janssen, IQVIA, Lundbeck, Mylan, Recordati, Sage, and Schwabe; and he has served on speakers bureaus for Abbott, Angelini, Aspen Farmaceutica S.A., Biogen, Janssen, Lundbeck, Recordati, Sage, Sanofi, Schwabe, Servier, Sun Pharma, and Vifor.
Ethics statement
Our research complies with internationally accepted standards for research practice and reporting. The present investigation was carried out in accordance with the latest version of the Declaration of Helsinki. The study design was reviewed by appropriate ethical committees. Written informed consent of all the participants was obtained after the nature of the procedures had been fully explained.
Open access
Comments
No Comments have been published for this article.