Borderline personality disorder and depression severity predict suicidal outcomes: A six‐month prospective cohort study of depression, bipolar depression, and borderline personality disorder

Suicide risk is high in patients with major depressive disorder (MDD), bipolar disorder (BD) and borderline personality disorder (BPD). Whether risk levels of and risk factors for suicidal ideation (SI) and suicide attempts (SA) are similar or different in these disorders remains unclear, as few directly comparative studies exist. The relationship of short‐term changes in depression severity and SI is underinvestigated, and might differ across groups, for example, between BPD and non‐BPD patients.


| INTRODUCTION
Suicide death (SD) is, in psychological autopsy studies, usually preceded by psychiatric morbidity, most commonly depression. 1,2 In epidemiological and cohort studies, major depressive disorder (MDD) and bipolar disorder (BD) confer suicide risk far exceeding the general population. [3][4][5] Borderline personality disorder (BPD) also confers a high risk of SD, 6 as much as 5%-10% in long-term follow-up. 7,8 Personality disorders, such as BPD, increase suicide risk, especially when complicated by depression or substance abuse. [9][10][11] Patients with major depressive episodes (MDEs) in MDD (MDE/MDD), BD (MDE/BD) and MDE with comorbid BPD (MDE/BPD) are thus three central patient groups with high suicide risk. However, differences in risks and specificity of risk factors between these groups are poorly known because of heterogeneity of previous studies in terms of setting and design, and a paucity of prospective studies.
Suicide death is one possible end result of a dynamic process in which suicidal ideation (SI) may fluctuate, potentially progressing to suicide attempts (SA) or SD. 12, 13 In mood disorder patients, SI seems to be almost a sine qua non for SAs. 14,15 Understanding which, when and why individuals proceed from ideation to action remains a challenge. In suicidology, this process has been scientifically explained by theories forming the so-called ideationto-action framework, which all emphasise that causes of SI may be different from those for action. 13 Thus, risk factors for SI, SAs and SD are unlikely to be identical, but may vary across different stages of the process. 16,17 Risk factors for SI, SAs and SD in MDD, BD and BPD have mostly been studied separately. In a recent meta-analysis, severe depression and negative psychosocial events significantly predicted SI and SAs, whereas severe depression and substance abuse of SD in MDD. 17 In BD, risk of SA is associated with episode polarity, comorbidities such as anxiety, substance use, and cluster B personality disorders, substance use (except cannabis), family history positive for SD, female sex and earlier age at onset, whereas SD was predicted by male sex and family history of suicide. 18 A study comparing suicidal behaviour in MDD and BD found trends towards differences in risk factors between BD and MDD patients, such as time spent depressed or otherwise ill and alcohol use being more relevant in BD and treatment-related aspects in MDD, but their statistical or clinical significance was not explicated. 19 In another comparative study finding BD patients to have a higher lifetime prevalence of SA than MDD patients, incidence was increased during illness episodes (in which BD patients spent more time), but was similar across groups during clinically similar episodes. 20 Comorbid BPD or BPD features increase the prevalence of SAs and severity of SI in patients with mood disorders in retrospective and cross-sectional studies. [21][22][23] The effects of personality disorders on SA risk in depression seems partly mediated by increased time spent depressed. 24 In a prospective study of inpatients with MDD, BPD and both, SAs were predicted by impulsivity and hopelessness in all three groups. 25 Hopelessness and depression are correlated to an increased risk of SI, SAs and SD. 26 Hopelessness has stateand trait-like qualities, 27 and covaries with depression severity, which partly explains the covariation with SAs and SI. 28 Overall, the differences in observed risk factors for suicidal thoughts and acts between MDD, BD and BPD likely reflect both differences in study methodology as well as genuine differences. To clarify findings, comparability of measures is needed.
Many studies have reported on the relationships between different BPD criteria and suicidal outcomes. 22,29,30 In ecological momentary assessment studies, SI in BPD was linked to interpersonal difficulties and

Significant outcomes
• Baseline severity of BPD features and hopelessness predict future risk of suicide attempts and significant suicidal ideation in patients with depression. • In biweekly follow-up with the modified PHQ-9, severity of depressive symptoms consistently covaried with the presence of suicidal ideation in unipolar, bipolar or BPD patients with depression. • Dimensional assessment of severity of BPD features and prospective monitoring of depression severity seem useful for assessing suicide risk in depression.

Limitations
• The 6-month follow-up may affect interdiagnosis differences and preclude drawing conclusions of long-term outcomes. • The study population consisted of psychiatric outpatients, which influences the generalisability of findings to other settings. • In the biweekly follow-up, patients were queried for thoughts of self-harm using an item of the modified PHQ-9 which included but was not fully specific for suicidal ideation, and future studies might more accurately focus on suicidal intent.
affective lability, 31,32 with negative affectivity serving as a possible link between external triggers and SI. The importance of fluctuating depressive symptoms as risk factors for SI and SAs in BPD patients remains unclear. We have previously reported that dimensionally assessed BPD severity correlates cross-sectionally with risk of SA and severity of SI in depressive patients, also those without BPD, 21 but the predictive significance of BPD feature severity remains unknown.
To optimise treatments targeted to reduce risk of SD and SA, identification of at-risk individuals is not enough-periods of increased risk are also important, since shortening their duration may ameliorate total risk. 12 Many risk factors for SI and SAs are timeinvariable (e.g., family history) 17 and help to identify atrisk individuals but not high risk periods. Furthermore, the risk factors for worsening SI in the short term (days to weeks), including basic relationships between the severity of SI and of depression, have rarely been investigated. 33 Among MDD patients with SI followed weekly, a decline in severity of depression and hopelessness preceded remission of SI, 34 but studies in MDE patients with BD or BPD are lacking. Although severity of depression is widely clinically used as a warning sign for increased suicide risk, the evidence for this practice is mostly crosssectional, 33 and longitudinal studies of the covariation of depression severity and suicidal thinking are needed.

| Aims
We investigated whether there are differences in SAs and SI in a prospectively followed cohort of MDE patients, depending on whether they suffered from MDD or BD, and whether they had comorbid BPD. We (1) hypothesised MDE/BPD patients to be at highest risk, followed by MDE/BD and MDD without BPD. We also investigated risk factors for prospectively observed significant SI and SAs, (2) hypothesising that BPD and severity of BPD features increase these risks even after adjusting for other relevant risk factors. Finally, we (3) hypothesised significant within-patient covariation between thoughts of suicide or self-harm and severity of depression in all three diagnostic groups.

| Recruitment and inclusion
This is a prospective cohort study of psychiatric patients with major depressive episodes. We have previously described patient recruitment in more detail. 21,35,36 In brief, we recruited treatment-seeking 18-50-year-old outpatients referred to secondary-level psychiatric services of the City of Helsinki, Finland, for depression. Recruitment was based on stratified sampling into three strata according to probable principal clinical diagnosis (MDE/MDD, MDE/BD or MDE/BPD) based on the referral and sex of the patient. Stratification allowed for enriching patients otherwise underrepresented, with an aim to recruit a sufficient number of participants with MDE/MDD, MDE/BD and MDE/BPD of both sexes. 35 We called possible recruits over the telephone, and, if they gave preliminary consent, we (J. L. S. and J. J. S.) met with them for diagnostic interviews. Consenting individuals were then interviewed with the structured clinical interview for DSM disorders (SCID I and II) and the Montgomery-Åsberg depression rating scale (MADRS). [37][38][39] Those meeting the inclusion criteria of current major depressive episode (MDE) and a MADRS score of 15 or more, but not meeting exclusion criteria (active substance use, psychotic illness, schizotypal or antisocial personality disorder, neurocognitive and/or sensory issues affecting study participation and insufficient Finnish language proficiency) were included in the study. We videotaped diagnostic interviews for testing of inter-rater reliability, which was found excellent, with Cohen's kappa values of 1.0 for MDD, 0.898 for BD, and 0.891 for BPD. 21,35 Our research protocol has been approved by the Ethics Committee of the Helsinki and Uusimaa Hospital District, and granted a research permit by the City of Helsinki.

| Study population and subcohort assignment
Of 1655 potential patients referred, 436 did not meet inclusion criteria, 855 were not contacted due to limited researcher capacity and 209 were not willing to participate in the study. We interviewed 155 patients, of whom 31 were excluded, and our initial sample was thus 124 participants. 21,35 We divided our participants into three subcohorts: (1) MDE in major depressive disorder (cohort inclusion criterion was no diagnosis of BD or BPD; MDE/MDD), (2) MDE in bipolar disorder (diagnosis of BD; MDE/BD) and MDE with comorbid borderline personality disorder (diagnosis of BPD; MDE/BPD). Participants meeting the criteria for both BD and BPD were assigned to the BD subcohort if they had type I BD, otherwise subcohort assignment was judged by researchers (J. L. S. or J. J. S.) based on presenting symptomatology. Unclear cases were discussed in the study group and a consensus decision about subcohort assignment was made.

| Interviews and questionnaires
At baseline, in addition to the SCID and MADRS, we investigated previous lifetime and recent SI and suicidal behaviour with the Columbia-Suicide Severity Rating Scale (CSSRS) 40 ; these results have been previously reported. 21 This instrument registers both SAs and other suicidal behaviours, such as interrupted attempts or preparations for attempts, during a defined period (in this case, the duration of this study). The CSSRS also quantifies the severity of SI on the suicidal ideation severity scale (SIS), and the intensity with the suicidal ideation intensity rating (SIIR). We furthermore measured the severity of BPD features or symptoms dimensionally at baseline with the borderline personality disorder severity index (BPDSI), 41 an investigator-rated interview-based instrument quantifying the frequency and severity of BPD-related symptoms and behaviours during the preceding 3 months. We rated manic symptoms at baseline with the Young mania rating scale (YMRS). 42 Patients were also asked to complete questionnaires, such as the hopelessness scale (HS), 43 the Beck depression inventory II (BDI-II), 44 the alcohol use disorder identification test (AUDIT), 45 and the overall anxiety severity and impairment scale (OASIS). 46

| Follow-up
After initial assessment, there was a follow-up period of at least 6 months for each subject. During this time, they were asked to complete a biweekly online questionnaire charting their depressive symptoms with the Patient Health Questionnaire 9 (PHQ-9). 47 We modified the PHQ-9 item 9 by splitting thoughts of death versus suicidal and self-harm ideation into two separate questions; the latter was used in analyses of SI. After a minimum of 6 months, we met the patients again for follow-up interviews, including a repeat CSSRS for the study period. Of 124 original study participants, 83 completed the followup CSSRS and 83 participated in the PHQ-9 follow-up, with a mean of 8.57 (standard deviation [SD] 2.76) completed questionnaires per participant, altogether there were 711 follow-up questionnaires answered.

| Data management and statistical analysis
We constructed a database with SQLite and analysed the data in R, using the packages lme4, pscl and glmmTMB. For significance testing of within group differences, we used Fisher's exact test and analysis of variance tests as appropriate. We analysed prospective risk factors for having made a SA and having clinically significant SI (defined as a SIS of 3 or higher, which signifies SI with some intent to act) using logistic regression models.
We also examined the concurrent relationships between depressive symptomatology in the biweekly online questionnaires (as indicated by the total PHQ-9 score for the previous 2 weeks with the self-harm item subtracted) and the risk for reporting thoughts of death, suicide or selfharm. We did this first through dividing the material into quintiles based on the total PHQ-9 score (without the selfharm item), and, second, by multilevel logistic regression models. Since long-term and short-term effects of an independent variable on a dependent variable are not identical (in extreme cases they may even be in opposite directions), 48 we aimed to disentangle longer term (i.e., over the study period, trait-like) depression severity from that of more temporary (state-like) severity variation with respect to their effects on the risk of self-harm ideation. For this reason, we constructed two predictive variables: the individual-level mean PHQ-9 score over the whole follow-up period, and the within-person meancentred PHQ-9 score for a particular follow-up instance (excluding the self-harm item from both).

| RESULTS
Characteristics of the study population as a whole and by subcohort are reported in Table 1.
The severity and intensity of SI as well as frequencies of suicidal and self-injurious behaviour during the study period are reported in Table 2. There were no suicide deaths during the study period.

| Suicide attempts
Since there were bipolar and borderline subjects in two of our three subcohorts, we also examined the odds of having suicidal behaviour when grouped by diagnostic status instead of subcohort. Out of 18 participants with a BPD diagnosis, 4 (22.2%) attempted suicide during the study period, while only 3 (4.23%) of the 66 others did so, this difference was statistically significant (Fisher's exact test p = 0.02867, odds ratio [OR] = 6.28, 95% CI 0.95-47.75; Kaplan-Meier survival analysis in Supporting Information).There were no significant differences in SA incidence between bipolar (5 of 39; 12.8%) and nonbipolar patients (

| Predicting clinically significant suicidal ideation
We examined a logistic regression model for clinically significant SI (defined as SIS ≥3) during follow-up, using age, sex, bipolar diagnosis and the MADRS, HS and BPDSI scores at baseline as predictors, also controlling for clinically significant SI in the past. Compared to a null model, this model explained the log odds with a significantly better fit (χ 2 (7) = 33.29932, p < 0.001), and the McFadden R 2 of the model was 0.312. The ORs of predictors are visualised in Figure 1 and a regression table including uncontrolled OR for each variable is given in Table 4. Regression models of the BPDSI subitems as predictors for SA and SIS severity are presented in the Supporting Information.

| Covariation of depression severity and thoughts of suicide or self-harm during follow-up
When dividing the PHQ-9 questionnaires into quintiles according to total PHQ-9 score, there were significant differences between the quintiles in the prevalence of SI (as per the modified item 9) at that time point (χ 2 (1) = 219.2, p < 0.0001). These proportions (for the whole cohort and grouped by BPD and BD diagnostic status) are depicted in Figure 2.
We examined this relationship through multilevel statistical modelling. First, we examined null models, void of predictors. In a one-level generalised linear (logistic) model, the Akaike Information Criterion (AIC) was 974.15 and the Bayesian Information Criterion (BIC) 978.76, whereas in a two-level model having a random intercept for individual mean level, the AIC was 565.02 and the BIC 574.26 indicating the superiority of the twolevel model nested within individuals.
We then added a within-person-centred PHQ-9 score from a specific time point, as well as the person-level PHQ-9 mean score as fixed-effect predictors for simultaneous SI. We found that this model fit the data better than the model without these predictors (AIC 483.2, BIC 501.7), and that there were highly significant associations between the risk of self-harm ideation and both the   18), but not the BIC (508.81); the regression coefficient between BPD diagnosis and SI risk was significant (p = 0.0161), but the interaction terms were not (BPD diagnosis Â centred PHQ-9 score p = 0.3023; BPD diagnosis Â PHQ-9 personlevel mean: p = 0.1135). Adding the BD diagnostic status to the original model yielded similar results (AIC 479.60, BIC 511.93); the bipolar diagnosis regression coefficient was not significant (p = 0.1438), but the interaction term between the centred PHQ-9 score and bipolar diagnostic status was (estimate 0.29479, p = 0.0120; total OR for centred PHQ score for bipolar patients 1.707). This indicates that the effect of depression severity on risk of simultaneous SI was highly significant in all patient groups, but that the correlation between changes in depression severity and risk of SI may be slightly stronger in BD than in non-BD patients.

| DISCUSSION
In this prospective cohort study of depressed outpatients, we found patients with BPD more likely to attempt suicide. We found severity of hopelessness and BPD features to independently predict the risk of clinically significant SI and risk of SAs. Unexpectedly, BD was associated with a lower risk. Furthermore, we found that during the biweekly follow-up, depression severity and SI covaried intra-individually in all three patient groups.

| Predictors for suicidal thoughts and attempts
Our findings support earlier findings of the central role of comorbid BPD as an important risk factor for SAs in depression, which markedly increases their clinical relevance. 21,22 Contrary to our hypotheses, we did not find that bipolar patients had a significantly increased risk of SA compared to MDD patients. Our findings highlight the role of severity of BPD features as a risk factor for significant SI and SAs in depression. Besides categorical BPD diagnosis, assessing severity dimensionally seems useful. Our findings regarding hopelessness as a risk factor for SAs and significant SI complement earlier findings, 26,28 and was seen in all three groups: MDD, BD and BPD patients.
Finding bipolarity to be associated with the lowest risk for significant SI is in seeming contrast to many earlier studies. 5,20 However, the risks of SI and SA are not identical and might vary over different timescales. Indeed, we previously reported that lifetime SA history was more common in BD than in MDD patients in this cohort. 21 The finding might also be a consequence of referral patterns specific to the Finnish public healthcare F I G U R E 2 Relationship between risk of suicidal ideation and simultaneous depression severity by quintiles. system, as the likelihood for referral to psychiatric care is likely higher for non-suicidal bipolar than non-suicidal unipolar depression.

| Covariation between concurrent depression severity and suicidal thinking
Another important finding is that, when followed biweekly, the severity of depressive symptoms and SI covary consistently. This finding is concordant with previous cross-sectional and longitudinal studies 28,34,49 with longer time scales, and as such represents a further step towards validating the causal role of depressive syndromes for suicidal thinking and acts. 33 BPD diagnosis status did not significantly influence this relationship, which indicates that this within-patient mechanism is similar in both BPD and non-BPD patients, or put differently, that worsening of depression seems to be a valid risk factor for SI in BPD as well as in non-BPD patients. This correlation may be slightly stronger in BD than in non-BD patients. Covariation with hopelessness partly explains the role of depression in SI, 28 but the precise role of other depressive cognitive biases 50 remains to be further studied.

| Strengths and limitations
One strength of this study is the clinically representative sample of three important patient groups in a directly comparative design, which permits the direct comparison of risk levels of SAs and SI, and risk factors for these, across depression patients with MDD, BD and/or BPD. The prospective design improves the scientific validity and clinical utility of our findings. Dimensional assessment of BPD severity and suicidal behaviour and ideation may increase validity since distributions of psychiatric symptoms in the general population tend to be continuous rather than discrete or bimodal, and there is likely to be a dose-response relationship between risk factors and outcomes. Additionally, the biweekly follow-up allowed us to investigate the ongoing relationship between self-harm ideation and depression severity over the shorter term using valid, within-person-centred, multilevel modelling, studying both structural and dynamic relationships and disentangling these from one another, with direct clinical implications for the use of worsened depression as a warning sign for suicide risk.
Our study had several limitations. First, the longitudinal time frame was designed to cover time to remission from a MDE. Our findings do not inform about longer time scales, which might also differ between diagnostic groups. Second, the moderate sample size increases the risk of inferential errors, but did not preclude us from reaching significant results. Third, we investigated psychiatric care outpatients using a stratified sampling approach, and validation of our findings in other populations are needed. Fourth, in the longitudinal follow-up we used a modified version of the PHQ-9 item 9 to differentiate passive thoughts of death from suicidal thinking, but even this did not differentiate between thoughts of self-harm from true suicidal ideation. This weakens the specificity of our findings. To put our findings in perspective, we found acts of nonsuicidal self-injury were reported by 10% of our cohort and 24% of the BPD subcohort during the study period, and 28% (all) and 64% (BPD subcohort) during their lifetimes. 21 Finally, this study was naturalistic and possible confounding effects of differing treatments on the outcomes remain unknown.
To conclude, in prospectively followed MDE, BPD patients were at higher risk of SAs than others. Severity of BPD features may be a relevant predictor for risk of significant SI and SAs for all patients with MDEs. Changes in depression severity may indicate a change in emergence, severity or alleviation of SI irrespective of diagnostic subgroup. BPD status at baseline and ongoing monitoring of changes in depression severity seem useful predictors for assessing risk of SAs both clinically and in research.

FUNDING INFORMATION
We acknowledge the kind support of Helsinki and Uusimaa Hospital District and Finska Läkaresällskapet. THR was supported by the Academy of Finland (grants 334057 and 335901 to THR).

CONFLICT OF INTERESTS STATEMENT
The authors declare that they have no conflicts of interest.

DATA AVAILABILITY STATEMENT
Data are not available due to restrictions imposed by the research permit and the Finnish data protection legislation.

ETHICS STATEMENT
The study protocol was approved by the Ethics Committee of the Helsinki and Uusimaa Hospital District and