Excess deaths in treatment-resistant depression

Background: Patients with treatment-resistant depression (TRD) have an increased mortality risk compared with other patients with depression, but it is not known how this translates into absolute numbers of excess deaths. Methods: Swedish national registers were used to identify a cohort of 118,774 antidepressant initiators 18–69 years old with a depression diagnosis. Patients who initiated a third consecutive treatment trial were classified as having TRD. Flexible parametric survival models were used to estimate the mortality risk due to all causes and external causes (suicides and accidents), comparing TRD patients with patients with other depression while adjusting for clinical and sociodemographic covariates and including interactions with TRD, age, and Charlson comorbidity index (CCI) for a number of somatic comorbidities. Standardized survival was estimated, as were numbers of excess deaths among TRD patients within each age and comorbidity category. Results: Compared with the mortality risk of other depressed patients, patients with TRD experienced excess deaths in most age and comorbidity categories in the range of 7–16 deaths per 1000 patients during 5 years. Highest numbers for all-cause excess deaths were found among patients 18–29 years old with CCI 1, where 16 [95% confidence interval 5–28] of the expected 37 [25–48] deaths per 1000 patients were excess deaths. The majority of the excess deaths were due to external causes. Conclusion: Patients with TRD experience significant numbers of excess deaths compared with other patients with depression.

actual impact that a condition, or preventive measure, may have on mortality. In our previous study, HRs comparing mortality among the TRD group with other depressed patients were higher in younger age groups and in groups without comorbidities. 9 However, a high relative effect may not necessarily mean that the absolute effect is highest in these groups.
An alternative way of illustrating mortality differences is the number of deaths that count as excess deaths within a specific time period (e.g., 5 years) in a group of patients, compared with that group if it were to experience the same mortality as a comparison group. This may also be a relevant measure for patients as well as for clinicians and decision makers. 10 Attempts to quantify the excess mortality from depression as number of deaths have been made, for example, in a recent study where 3.5% of deaths in the general population in the United States (US) were deemed attributable to anxiety and depression 11 ; however, this measure does not convey excess deaths within the depressed population and does not take excess deaths in treatment-resistant patients into account.
Building on our previous investigation of the relative mortality risk associated with TRD, the purpose of this study was to estimate the absolute effect, that is, numbers of excess deaths among patients with TRD compared with other patients with depression, by using a cohort of patients with depression identified in nationwide Swedish registers. This is presented by age-and comorbidity groups.

Material and methods
Data sources and study population As previously described, 9 we used a combination of Swedish national registers to identify the study cohort, covariates, and outcomes. All patients 18-69 years old who filled a prescription for an antidepressant drug (any drug with an ATC-code starting with N06A) in Sweden between 1 July 2006 and 31 December 2014, without any previous antidepressant prescription for 180 days, were identified through The Prescribed Drug Register (PDR), 12 which contains data on all prescribed and dispensed prescriptions in Sweden since 1 July 2005. Among these, patients who had a diagnosis of depression (ICD-10 F32, F33, or F34) in the National Patient Register (NPR) within a time interval of 30 days before, and up to 365 days after, the filled prescription were included in the depression cohort. 13 The NPR contains data on all diagnoses and procedures during in-patient care and specialized out-patient care -excluding primary care -according to the International Statistical Classification of Diagnoses, 7th-10th editions (ICD 7-10) with complete national coverage for in-patient care since 1987, and for outpatient care since 2001. Exclusion criteria were fills of lithium (N05AN01), antipsychotics (N05A, except for lithium), and the anticonvulsants valproate (N03AG01), lamotrigine (N03AX09), and carbamazepine (N03AF01) during 180 days before the first antidepressant prescription, and also those with procedure codes for electroconvulsive therapy (ECT; DA006, DA024-25) or repetitive transcranial magnetic stimulation (rTMS; DA022, DU050). Patients with a previous history of psychosis (ICD-10 F20-F29), mania (F30), bipolar disorder (F31), or dementia (F00-F03) were also excluded, as were patients who were not residents in Sweden for the full 180 days according to the Total Population Register before the first antidepressant prescription filling. 14 The index date was defined as the date when both the criteria of a prescription and a diagnosis were fulfilled. Patients could enter the study at any time during the study period when inclusion criteria were met.

Definition of TRD
Patients were classified with TRD from the starting date of a third treatment trial for depression after having undergone two adequate trials. An adequate treatment trial was defined as lasting for at least 28 days. Subsequent treatment trials after the first dispensed antidepressant trial had to include at least one fill of a different antidepressant (as monotherapy or in combination with the previous one), add-on medication to the antidepressant (antipsychotics or anticonvulsants), or administration of ECT or rTMS. Durations of trials were estimated by semi-manually reading prescription dosage texts and taking into account the number of tablets dispensed at each fill, and by action codes for ECT/rTMS in the NPR. Treatment gaps of >28 days were not allowed in order to emulate consecutive treatment trials. Distribution of medication/treatments used for the third treatment trial is shown in Supplemental Table S1.

Outcomes
The main outcomes were all-cause mortality and death from external causes (including suicides journals.sagepub.com/home/tpp 3 and accidents; ICD codes V01-Y98) as registered in the Cause of Death Register. 15

Estimation of hazard rates and HRs
Flexible parametric models (FPM) were used to model mortality in the cohort. 18,19 A FPM can be used to estimate HRs, similarly to Cox regression, but, unlike the Cox model, it also estimates the baseline hazard rate using restricted cubic splines. Two FPMs were fitted first with all deaths considered as events, and secondly with death from external causes (ICD codes V01-Y98) as events, censoring at other deaths. Patients were followed from the index date until death and censored at emigration, end of study, or if any of the exclusion criteria were fulfilled. The exposure of interest, TRD, was treated as a time-varying covariate where a patient moved to the TRD group (i.e., from unexposed to exposed group) when the criteria for TRD were fulfilled. The two FPMs also adjusted for age at index date (using splines), sex, previous self-harm, previous substance use disorder, previous depression, other psychiatric comorbidity, and CCI (categorized as 0, 1, ⩾2 in the model for death due to external causes), and all of these were allowed to have time-varying effects (non-proportional hazards). Interactions between TRD and age (in categories 18-29, 30-49, 50-69), as well as TRD and CCI were included, since those were the interaction effects shown to be important in the previous study. 9 Both FPMs had four degrees of freedom for modelling the baseline and two for time-varying effects.

Estimation of standardized survival and excess deaths
Based on a fitted FPM, the survival probability at a certain point in time can be estimated for each patient, given the values of all variables included in the model for that specific patient. By taking the average of the estimated survival probabilities across all patients, a marginal survival probability is estimated, internally standardized to the covariate distribution within the cohort. Marginal allcause 5-year survival was estimated for the group that developed TRD, separately for each combination of age group and CCI category. Within the cohort, TRD is a time-varying exposure; however, in the estimation of marginal survival, TRDstatus was assumed to be fixed, so the survival was estimated as if everyone was TRD from start of and throughout 5 years of follow up. In the actual cohort, not all patients had 5 years of potential follow up, and even those who did, might not have 5 years of potential follow up as TRD exposed since TRD varies over time.
When interest lies in comparing survival between groups, while adjusting for other covariates, standardized survival can be used, by standardizing over the same covariate distribution for all groups. The average 5-year survival was therefore also estimated for non-TRD, separately for each combination of age group and CCI category, but standardized to the covariate distribution of the TRD patients. Since the only difference between the two standardized estimates within each age group and CCI category is the value set for the exposure of interest, the standardized survival estimates are comparable in terms of the confounder variables included in the model, and can be viewed as adjusted survival estimates.
Based on the 5-year standardized survival estimates, the expected number of deaths within 5 years was estimated, within each age group and CCI category. We then estimated the difference in number of expected deaths within 5 years for each age and CCI group combination, from the two standardized survival estimates. This can be interpreted as the number of deaths that could be avoided within 5 years among TRD patients, if they had the same mortality rate as patients with other depression (given their covariate distribution), within that specific age and CCI group. To aid comparison across age and CCI groups, we also estimated the number of expected deaths within 5 years for 1000 patients of each group. However, for each age and CCI group, the marginal/standardized survival was estimated given the distribution of other covariates within each of the groups. Differences in the number of expected deaths, between age and CCI groups, can therefore be due, to some extent, to differences in the distribution of all other factors taken into account in the analysis. The comparison between TRD and non-TRD is, however, due to the standardization, adjusted for all covariates. Standardized survival and number of excess deaths were also estimated for deaths due to external causes, using the same approach as described above.

Ethical permission
The study was approved by the regional ethical review board in Stockholm (no. 2017/1236-31/2).

Results
The In Table 3, the same estimates as above are presented for deaths from external causes only. Again, the highest HRs and absolute numbers of estimated deaths were seen in the numerically largest patient categories. Numbers of estimated excess deaths from external causes were similar to estimates for all-cause deaths among younger and physically healthy patients -for instance, the same estimate, 10 deaths (95% CI 5; 15) per 1000 patients within 5 years was found among patients 18-29 years old with CCI 0 in both analyses -suggesting that the absolute majority of all expected deaths among these patients were from external causes. When estimating the numbers of excess deaths from external causes, the numbers of significant excess deaths were of a similar magnitude as for all-cause deaths, ranging from 9 to 15 per 1000 patients with TRD survival within 5 years. The largest number was estimated at 15 (95% CI 2; 28) of a total of 34 (95% CI 22; 47) expected TRD deaths in the age group 50-69  Deaths from external causes, i.e., suicides and accidents (ICD-10 codes V01-Y98).
with CCI 1. For patients with CCI 2 and above, the point estimates for excess deaths were lower, but statistically non-significant.

Discussion
This is the first study to translate the increased mortality in TRD compared with other depression into estimated numbers of excess deaths. Excess deaths were found in most age and comorbidity categories, ranging from 7 to 16 deaths per 1000 patients with TRD during 5 years. Numbers of excess deaths from external causes, i.e., suicides or accidents, were similar to numbers of excess all-cause deaths, implying that the majority of excess deaths in TRD are due to external causes. The results illustrate how excess deaths can be used to expand and complement the previously reported increased HRs for mortality, standardized mortality ratios, and excess mortality rate ratios among patients with TRD. 9,20,21 Interpretation The interactions of mortality with age and physical comorbidity were investigated specifically in this study. Within each age category, most point estimates for all-cause excess deaths were somewhat higher among patients with comorbidities (CCI ⩾1) compared with patients without comorbidities. This higher number of excess deaths could, if accurate, be attributed to several factors. First, patients with somatic comorbidity may be at higher risk for developing TRD compared with other patients, although evidence in the literature for this association is conflicting. 22 It is also possible that TRD itself may have a negative impact on the course of somatic disease, and therefore contribute to the excess deaths among these patients. Studies have reported an increased mortality after acute myocardial infarction among patients with TRD compared with other depression, 23 and persistent depressive symptoms may increase mortality after somatic hospitalization. 24 Also, the construction of the CCI means that patients with different severity of the same somatic disease are classified with an equal score. 16 This may leave residual confounding as severely ill patients with depression may be at higher risk for both TRD and death. 22 In the analysis including only deaths from external causes, no clear pattern regarding somatic comorbidity could be seen, apart from lower point estimates for excess deaths among patient with major comorbidities (CC1 ⩾2). This may be due to suicide and somatic disease being competing risks of death for these patients. Regarding the interaction with age, however, both suicide risk and actual numbers of suicide are known to increase with age in the general population. 10 In the present study, there were also slightly higher -albeit non-significant -numbers of excess deaths from external causes in 5 years per 1000 patients in the oldest age segment. However, as this was also the smallest group in numbers, absolute numbers of excess deaths were higher in the two younger age categories. Together, the investigation of these interactions emphasizes the importance of applying different measures of mortality in a population and identifying subgroups at risk, as patients who are in higher risk categories may not impact the overall numbers if numerically small.

Implications
Previous research has established that the allcause mortality among patients with depression is markedly higher than in the general population, and that this is even more pronounced for suicide. 25 The results of the present study regarding excess deaths in TRD may be of both clinical and societal importance. Rates of TRD among patients initiating treatment for depression have been estimated to be 6%-29% in administrative data studies, 5,6 and as high as 30%-50% in clinical studies. 3,4 Considering that life-time risk for depression in the general population has been estimated at 15%-20%, 26 the number of excess deaths in TRD is likely to be numerically impactful when transferring these rates to larger populations. According to the estimates in this study, TRD patients may experience up to twice as many deaths than if they had the survival of other patients with depression. Especially worrisome may be the estimated excess deaths among young, somatically healthy, patients, of which the large majority was due to external causes.
The patients with TRD in this study have all been administered several treatment trials for depression, as a proxy for depression that have not responded adequately to treatment, which may be a contributing factor to the excess deaths reported in this study. However, the group classified with TRD may be heterogeneous, as a substantial number of patients who present with TRD may have been misdiagnosed with depression and/or suffer from other undiagnosed psychiatric comorbidities, 27,28 which could lead to worse antidepressant treatment results and may contribute to the excess deaths seen here. The TRD rate of 13% found in this cohort is substantially lower than the 30%-50% found in clinical studies, 3,4 but is similar to other epidemiological studies in countries with comparable national databases (Denmark 14%, Taiwan 21%), 29,30 and also to using other algorithms for defining TRD in Swedish data (9%-19%). 31 The lower number of patients with TRD identified in observational studies compared with clinical studies may represent a lower tendency in reallife practice to administer more than two treatment trials, and some evidence points to the view that the smaller the proportion of patients identified with TRD, the higher the morbidity and poorer outcomes among patients. 31 Regardless of the underlying mechanism, TRD as defined in this study -starting a third treatment trial for depression -seems to increase the risk for mortality and the number of deaths.

Strengths and weaknesses
Strengths of this study include the cohort design with use of nationwide registers of high quality and virtually no loss of follow up, allowing for the calculation of survival rates in a large population for estimation of excess deaths. Among the limitations of this study is the use of a pharmacoepidemiological model for identifying TRD without access to clinical data aside from registered diagnoses. This lack of information includes unknown reasons for continuation, discontinuation or change of treatment trials (e.g., efficacy, non-efficacy, side effects, or lack of adherence). Patients over 70 years of age were not included due to hypothized competing risks between age-related deaths and deaths from sickness and external causes, and also due to likely differences in antidepressant treatment patterns compared with younger patients. Also, due to diagnoses registered in Swedish primary care not being included in the NPR, only patients seeking specialized psychiatric care were included. This may have attenuated the risk differences in the group comparison, for example, not including patients with milder or easily treated forms of depression. Finally, the PDR contains data from July 2005 onwards, which means that patients with TRD status before this date may have died before or during follow up without study inclusion.

Conclusion
Patients with TRD experience all-cause excess deaths in the range of 7-16 deaths per 1000 patients during 5 years compared with if they had similar mortality as other patients with depression. More knowledge is needed regarding to what extent more effective treatments for depression may lower the number of excess deaths.

Conflict of interest statement
PB and JR are affiliated with/employed at the Centre for Pharmacoepidemiology at Karolinska Institutet, which receives grants from several entities (pharmaceutical companies, regulatory authorities, contract research organizations) for the performance of drug safety and drug utilization studies. MN is an employee and stockholder of Janssen Inc. and contributed to conceptualization of the study and manuscript writing.

Data availability
The data used for this study may not, according to the ethical permission granted for its use, be shared by the authors to a third party. It is accessible by application to the Swedish authorities (The National Board of Health and Welfare and Statistics Sweden).