To be included in this study, participants had to present with a single or recurrent, unipolar MDD episode according to the DSM IV-TR. The participants were then selected if they had a HDRS₁₇ score > 20 after the failure of at least one antidepressant, delivered at an efficacious dosage for at least 6 weeks. The exclusion criteria were age under 18, other axis I disorders (except for anxiety disorders) including substance use disorder (except for nicotine), somatic or neurological disorders, failure to respond to venlafaxine during the current episode, pregnancy, previous rTMS, and rTMS contraindications.

The variables for the current analysis included: age, sex, education level, number of previous hospitalizations, age at onset of the first MDD episode, duration (months), total number of medications for the current episode, number of suicide attempt, presence of anxiety or personality disorder, family history of psychiatric disorder, and invalidity pension. Anxiety and personality disorders were declared by investigators following a clinical assessment by a psychiatrist with a lifetime perspective, based on the DSM IV-TR criteria and a structured interview using the MINI 5.0. BDI₁₃ was assessed at baseline and every week during the treatment course.

Qualitative and quantitative variables were described as number, percentage of each category and as mean and standard deviation (SD) respectively, each week, and for each treatment arm. The distribution of missing data over time in each group was estimated with a Kaplan-Meier estimator and compared by a log-rank test. The first step of the analysis consisted in investigating the factor structure (i.e., the number and composition of domains in terms of included items) of the French version of the BDI₁₃. Indeed, to our knowledge, the factor structure of the French version of the BDI₁₃ has not recently been investigated in such a large sample of depressed patients. Moreover, the validity of translated questionnaires has to be evaluated as language and cultural differences may depend on cultural context (Guillemin et al., 1993). The BDI₁₃ was translated into French (Delay et al., 1963) and while its psychometric properties (Collet and Cottraux, 1986; Cottraux, 1988) have already been assessed, it was more than 25 years ago on small samples of depressed patients (45 and 50 patients, respectively). Moreover, the number of domains of the BDI₁₃ was debated (Beck et al., 1988). Finally, ensuring the validity of the BDI₁₃ domain scores is a prerequisite for studying how they change over time. Hence, a confirmatory factor analysis was first performed on the 3-factor structure of the original version (Sad Mood, Guilt, and Performance Impairment). Good (acceptable) fit was indicated by the following criteria: root mean square error of approximation or RMSEA ≤ 0.05 (0.08), standardized root mean square residual or SRMR ≤ 0.05 (0.10) and comparative fit index or CFI ≥ 0.97 (0.95). In addition to fit assessment, reliability of the BDI₁₃ domains was assessed by internal consistency using the Cronbach's alpha coefficient, α. Domains were considered reliable if α > 0.70. The confirmatory factor analysis suggested that the French version of the BDI₁₃ does not satisfactorily confirm the original structure (RMSEA = 0.044, SRMR = 0.081, CFI = 0.936) and only the Performance Impairment domain was considered reliable as Cronbach alphas were equal to 0.62, 0.68, and 0.74 for Sad Mood, Guilt, and Performance Impairment domains, respectively. Hence, an exploratory factor analysis with orthogonal varimax rotation was performed.

The second step of analysis in this study consisted in fitting linear mixed models to explain changes in BDI₁₃ scores throughout treatment course and identify clinical variables associated with these changes. Separate linear mixed models were fitted on each domain score of BDI₁₃, identified in the first step of the analysis. Linear mixed models were chosen to analyze these longitudinal data as it is possible to characterize the mean change of the population over time with the fixed effects of the model while modeling individual variation around the mean trajectory of the score with random effects. The fixed effect part of the linear mixed model explaining each score included the treatment arm, the time, the interaction between time and treatment arm, and a set of covariates and their interaction with treatment arm at first. As little was known about the covariates associated with BDI₁₃, the set of covariates was defined as the variables listed in Variables section and associated in univariate analyses with changes in BDI₁₃ domain scores over time (p < 0.20). For each BDI₁₃ score separately, the set of covariates in the fixed effect part of the linear mixed models was refined by selecting variables by backward elimination (p < 0.05). The final mixed model of BDI₁₃ scores estimates the associations between depression changes, treatment arm and the retained variables.

Level of significance was set at 5%. Psychometric analyses were performed with Stata 15 software, and other analyses used SAS software (version 9.4, NC, USA).

All patients gave consent accompanied by a comprehensive assessment form. The study was registered under the ClinicalTrials.gov identifier: NCT00714090, conducted in accordance with the Declaration of Helsinki and was approved by a local ethics committee (CPP Sud est 6 #AU732).

A structure with 3 domains showed acceptable fit (RMSEA = 0.037, SRMR = 0.064, CFI = 0.954). These domains were: Sad Mood (item 1: sadness, item 2: pessimism), Negative Self-Reference (item 3: sense of failure; item 5: guilt; item 6: self-hate; item 7: suicidal ideas; item 10: negative body-image), and Performance Impairment (item 4: dissatisfaction; item 8: social withdrawal; item 9: indecision; item 11: working difficulties; item 12: fatigability). Negative Self-Reference and Performance Impairment domains showed acceptable reliability with Cronbach alpha coefficients equal to 0.73 and 0.72 respectively, contrary to the Sad Mood domain (α = 0.55).

As previously reported by Brunelin et al. (2014), the patient population (mean age of 54 +/-11 (SD) years) exhibited characteristics of severe and difficult-to-treat depression, as shown in Table 1, with: a mean HDRS₁₇ 25.86 +/- 3.7; treatment resistance (>2 treatment lines on average); several previous hospitalizations (3 on average); family history of mood disorders (59%); suicidality [about 41% committed at least one suicide attempt(s)]; long lasting episode duration (mean: 19 months); severe functional impairment (48.5% had long term illness exemption); and comorbidities, especially anxiety (23.8%) or personality (30.5%) disorders. Tolerability was good and premature withdrawal due to side effects were observed in 11.1% cases in rTMS, venlafaxine and combination, overall. The distribution of missing data was not significantly different between groups (p = 0.59).

After 6 weeks of treatment, we observed in the whole sample a significant decrease of depression scores with a mean decrease of−0.6,−0.3,−0.6 points per week on average for the Negative Self-Reference, Sad Mood, and Performance Impairment domains of the BDI₁₃, respectively.

Treatment groups were not significantly associated with changes in any domains of the BDI₁₃. Patients' characteristics significantly associated with higher mean scores of Negative Self-Reference at baseline and during treatment course were: personality disorder (+2.1 points) or past history of suicide attempt (+1.7 points), age under 65 years (+1.5) or female (+ 1.1 points) as presented in Table 2.

Specifically, the mean score of the Negative Self-Reference domain was the highest at baseline and also during treatment course for women under 65 years with personality disorder and a history of suicide attempt(s) as compared to other patients. The effects of the most significant covariates, namely personality disorder, past history of suicide attempt and age, on the estimated means of Negative Self-Reference score over time from the linear mixed model are represented on Figure 1. On Figure 1A, women under 65 years old with no personality disorder nor history of suicide attempt have a mean score of Negative Self-Reference of 5.6 points (intercept estimate, Table 2) at baseline and of 2 points after 6 weeks of treatment (decrease of depression:−0.6 points per week, p < 0.05) as represented with a solid black line. The decrease of depression is similar for women under 65 years old with personality disorder and a history of suicide attempt (Figure 1A, dotted gray line) but they have on average higher scores of Negative Self-Reference at baseline, i.e., 9.4 points (intercept +2.1 points for personality disorder and +1.7 points for history of suicide attempt, Table 2) and after 6 weeks of treatment, i.e., 5.8 points. Hence, their levels of Negative Self-Reference are higher at baseline and also remain higher during the 6-week follow-up. The same trends are observed for patients aged 65 years or older (Figure 1B) with significantly lower mean scores (age estimate:−1.5 points, Table 2). The effect of age can be observed by comparing the curves with the same layout between Figures 1A,B, for example the solid black lines for women with no personality disorder nor history of suicide attempt. None of the variables, except time, were associated with Performance Impairment and Sad Mood domain scores at baseline and over time.

To the best of our knowledge, this is the first study focusing on longitudinal changes in self-reported depressive domains including negative self-reference and identifying clinical variables associated with these changes. It was a secondary analysis of the largest clinical trial including both rTMS and antidepressant treatments (Brunelin et al., 2014).

First, we described a 3-factor structure of the BDI₁₃ (i.e., Negative Self-Reference, Sad Mood and Performance Impairment) showing acceptable fit. More specifically, Negative Self-Reference mainly dealt with negative self-regard (feeling of failure; guilt; self-disgust; negative self-image; suicidal tendencies). Similar domains, i.e.: “Negative Attitude Toward Self” and “Performance Impairment,” were also reported for the BDI-II scale (Shafer, 2006; Wang and Gorenstein, 2013).

Second, previous suicide attempt(s) and personality disorders were associated with higher depression scores at baseline and during treatment course for the Negative Self-Reference domain of the BDI₁₃, regardless of treatment arm.

These higher scores are clinically meaningful considering that full remission from the patient point of view require the restoration of self-confidence and a “normalization” of self-perception (Zimmerman et al., 2012). Moreover, Negative Self-Reference is a key prognostic dimension since resistance to antidepressants is associated with excessive self-focus, negative self-regard, cognitive inflexibility, hopelessness, feeling of inadequacy, and suicidal tendencies (Mor and Winquist, 2002; Nolen-Hoeksema et al., 2008; Zahn et al., 2015; Kneeland et al., 2016; Cândea and Szentagotai-Tătar, 2018; Huprich, 2019; Rush et al., 2019). A woman under 65 years old and suffering from personality disorder and past suicide attempt(s) history can have up to 6.4 additional points on Negative Self Reference domain compared to older male patients without these antecedents, so that this subgroup is more likely to present a sustained higher qualification of episode severity, for example from a mild (4-7) to moderate (8-15), or from moderate to severe (16 and above). Thus, monitoring negative self-regard during treatment course is an important clinical issue.

We observed that suicide attempters had higher level of negative self-reference, which is consistent with the literature. Suicidal behaviors and negative self-reference seem to have a bidirectional link: self-contempt may be a risk factor for suicidality and suicidal behaviors lead to negative attitude toward oneself (Jollant et al., 2017; Rüsch et al., 2019; Solano et al., 2019). Moreover, recent findings based on the assessment of functional connectivity between neuronal networks, showed particular patterns of activation of the Default-Mode Network (DMN) supporting self-referencing process, in suicide attempters vs. non-attempters suffering from MDD, and healthy controls. Authors suggested that suicidal behavior may be due in part to impaired self-referential thought processing and alters cognitive flexibility supported by the Central Executive Network (CEN) (Malhi et al., 2020). This is consistent with the description in MDD of an imbalance between an increased resting state activity in the DMN (self-focus) (Nejad et al., 2013), and a hypo-activity of the CEN (supporting executive functioning) (Boeker and Kraehenmann, 2018). In summary, our findings confirm the importance to measure self-reference domain in the follow-up of suicide attempters' follow up.

Personality disorders are known to be a risk factor for MDD (Bennabi et al., 2015; De Carlo et al., 2016; Kautzky et al., 2017; Roselli et al., 2020). They expose to negative affectivity and negative self-regard making it difficult to disentangle the two diagnoses during a depressive episode. Symptoms explained by the depressive state are more likely to change during treatment than those related to temperament or personality. Studies investigated the role of temperament (mainly using the Cloninger Temperament Inventory) and highlighted the importance of two domains -High Harm Avoidance (reflecting negative affectivity) and Low Self-Directedness- as trait markers associated with MDD (Zaninotto et al., 2016), as well as suicidality (Erić et al., 2017).

Trait malignant self-regard can explain higher level of depression and poorer outcome. However, little is known specifically about the links between personality disorders and self-reference process. Most of published studies focused on borderline disorders which is not the most frequent personality disorder associated with TRD. Moreover, physiopathology of negative self-referencing process has been little investigated, especially how the dorsomedial prefrontal cortex integrates increased attention to negative stimuli with an increased attention to the self in patients with personality disorders and MDD vs. MDD alone (Lemogne et al., 2011).

In summary, Negative Self-Reference may be related to both MDD state and personality traits, and patients with personality disorders are more likely to remain impaired on that clinical dimension.

Our findings highlight the value of focusing on domains rather than only on a global score given that depression may not be a homogeneous syndrome (Fried and Nesse, 2015). The impact of personality disorder and/or previous suicide attempts (representing 30.5 and 41% patients in our sample, respectively) on the Negative Self-Reference domain may draw the attention on the influence, from the patient's perspective, of those two variables on MDD prognosis whatever biological treatment type (brain stimulation and antidepressants). Our findings highlight the interest to consider interventions targeting negative self-regard in patients with personality disorder or past history of suicidal attempts. In line with this, Cognitive Behavioral Therapy may be efficacious for treating individuals with low self-esteem, according to changes in self-reported measures (Kolubinski et al., 2018). Several other psychotherapeutic approaches targeting self-representations and reappraisal, like psycho-education or cognitive remediation, are useful in addition to biological treatments (Pandarakalam, 2018).

Identifying such clinical factors associated with Negative Self-Reference pave the way for concomitant synergistic adjuvant interventions that may be promising throughout their ability to reduce ruminations, negative self-image, and improve cognitive capacities as well as perceived self-efficacy (Donse et al., 2018; Neacsiu et al., 2018; Laird et al., 2019; Wilkinson et al., 2019). Surprisingly, anxiety disorders were not significantly associated with the negative self-reference domain, despite its prevalence in TRD (Nuñez et al., 2018). We could have expected also higher negative self-referencing process in patients with baseline Generalized anxiety (GAD) since rumination, which frequently consists in negative self-judgment, is a core feature of both GAD and MDD (Merino et al., 2016). One can hypothesize that GAD was more associated with other features not captured by the BDI₁₃, such as worry about life circumstances, anxious apprehension of events or cognitive bias toward potential threat (Crocq, 2017). Contrariwise to our expectation that older age would be associated with higher depressive symptoms level, patients older than 65 years demonstrated lower Negative Self-Reference domain score. This could be explained by a positive cognitive bias and reappraisal of symptoms, which has been described in aging (Orth et al., 2010; Whitehead and Bergeman, 2016).