Participants
We used the nationwide population-based South Korean National Health Insurance Service-National Sample Cohort (NHIS-NSC) database [26], which contains the data of approximately 1 million people (2.2% of the total Korean population) and uses systematic stratified random sampling with proportional allocation within each stratum (age, sex, insurance eligibility status, and income level). As the National Health Insurance (NHI) program requires mandatory health insurance for all Korean citizens, the database is representative of the South Korean population. All insurance claims are catalogued in this system, and the medical information in the database is provided exclusively by healthcare providers. Every NHI member older than 40 years is eligible for biennial cardiovascular health panel screening and cancer screening for free or at a minimal cost depending on income [27]. Prior studies have validated the NHIS-NSC data for some chronic diseases, including stroke and dementia [28-31]. Yet, the validation of the International Classification of Diseases, 10th revision (ICD-10) codes for depression diagnosis has not been evaluated [32].
Of the 1,044,097 individuals who were enrolled in the cohort during the baseline period (January 1st, 2002 to December 31st, 2004), we included 244,920 individuals over the age of 50 years and then excluded 2,683 who had or received a dementia diagnosis during this period (Fig. 1). We followed up the participants from the index date (January 1st, 2005) until December 31st, 2013, or until the date of dementia onset within that period, defined as the follow-up period. Patient death and the end of the follow-up were treated as censoring events in the analyses. During the follow-up period of up to 9 years, individuals were classified into the following four groups based on their depression or CVD diagnosis: individuals with neither depression nor CVD, those with depression alone, those with CVD alone, and those with both depression and CVD.
The final analyses included a total of 242,237 participants. The institutional review board of the Samsung Medical Center, Seoul, South Korea approved this study. All data were anonymized and kept confidential, and thus, the need for obtaining participant informed consent was waived.
Exposure variables: depression and CVD
Depression was defined as the assignment of an ICD-10 code for depressive disorder (F32 or F33) and the documented administration of antidepressant medication (Additional file 1: Table S1) from the first day of depression diagnosis during the follow-up period as previously described [33-35]. CVD was defined as the assignment of an ICD-10 code for CVD (I60-69) as a primary diagnosis and two or more hospital visits during the follow-up period [35]. Unlike the medications used for depression or dementia, CVD medications are not exclusive to the disease. Therefore, we defined CVD based on the ICD codes and multiple hospital visits regardless of the medications prescribed.
Outcome of interest: dementia
The primary outcome was overall survival free of dementia. We defined dementia as the assignment of an ICD-10 code for dementia (F00-03, G30-31) and the documented administration of anti-dementia medication (donepezil, rivastigmine, galantamine, or memantine) during the follow-up period [33, 36, 37]. As mentioned above, we excluded all cases with a dementia diagnosis during the baseline period to focus on incident cases of dementia. Dementia subtype was set as the secondary outcome and was defined based on the initial ICD-10 code assigned on the first day of dementia diagnosis: F00 and G30 for Alzheimer's disease (AD); F01 for vascular dementia (VD); and F02-03 and G31 for other dementias (non-AD or non-VD).
Covariates
All analyses were adjusted for potential confounding variables—age, sex, residential area, income level, and comorbidities—during the baseline period from 2002 to 2004. With regard to residential area, the capital city of South Korea and the surrounding metropolitan cities (Seoul, Incheon, Gyeonggi-do) were designated as the ‘capital region’; all other regions were designated as ‘non-capital regions.’ Income levels were classified as low (up to the 30th percentile), middle (30th to 70th percentile), and high (70th to 100th percentile). The ICD-10 codes from the Charlson comorbidity index, which is used widely to adjust for the effects of comorbidities [38], were used. We defined the comorbidities based on the ICD-10 codes for each disease (Additional file 1: Table S2) and two or more hospital visits during the baseline period.
Interaction measurements
We assessed additive- and multiplicative-scale interaction measures to examine the interaction effect of depression and CVD on dementia onset. In terms of the additive interaction, we derived the attributable proportion due to interaction [AP; equation (1)], relative excess risk due to interaction [RERI; equation (2)], and synergy index [SI; equation (3)]. The AP is the proportion of the risk due to the interaction in the doubly exposed group (null hypothesis: AP=0). When RERI is positive, it indicates increased risk due to the additive interaction (null hypothesis: RERI=0). SI can be interpreted as the ratio of an increased risk due to both exposures to the sum of individual increased risks (null hypothesis: SI=1). These were used to assess whether the risk due to having both diseases is greater than the sum of the risks due to each condition [39, 40].
The multiplicative-scale interaction [equation (4)] has been widely used to examine interaction effects by identifying whether the risk due to having both diseases is greater than the product of the risks due to each disease alone (null hypothesis: multiplicative interaction=1) [39, 40].
In equations (1-4), HRDepression & CVD represents the hazard ratio (HR) for those who have both depression and CVD. HRDepression and HRCVD represent the HR for those who have depression or CVD alone, respectively, compared with individuals who have neither of the conditions.
Statistical analysis
We used Cox proportional hazards regression models to determine the adjusted hazard ratios (aHRs) and 95% confidence intervals (CIs) of depression, CVD, or comorbid depression and CVD for dementia incidence. As the cohort design can cause an immortal-time bias, we used a time-varying Cox regression model to prevent time-related biases [41, 42]. An unadjusted time-varying Cox regression analysis was performed (Model 1), followed by a demographic characteristics-adjusted (Model 2; adjusted for age, sex, residential area, and income level) and a comorbidity-adjusted (Model 3; adjusted for myocardial infarction, congestive heart failure, peripheral vascular disease, chronic pulmonary disease, connective tissue disorder, peptic ulcer, mild liver disease, uncomplicated diabetes, complicated diabetes, hemiplegia, moderate or severe renal diseases, non-metastatic solid cancer, moderate or severe liver diseases, and metastatic solid cancer in addition to the demographic characteristics in Model 2) analysis. The proportional hazards assumption was graphically tested and verified using the Schoenfeld residual method; no variables violated the assumption.
First, we used a log-rank test and evaluated independent associations of depression and CVD with subsequent dementia using aHRs and 95% CIs in two separate regression models. In this analysis, exposure of interest (depression or CVD) was treated as a time-varying variable, and the other comorbid illnesses were regarded as time-fixed confounders to be adjusted. Next, we examined the interaction effect of the two exposure diseases by calculating the additive (AP, RERI, and SI) and multiplicative interaction. We verified the significance of the interaction term and then stratified each subgroup based on age or sex. We then conducted a subgroup analysis using a fully-adjusted Cox regression model (Model 3; demographic characteristics and comorbidities adjusted) in which dementia subtypes (AD, VD, and non-AD or non-VD) were accounted into the outcome variables.
We also carried out sensitivity analyses to ensure the robustness of the results. First, lagged-time analysis was conducted because depression that occurs shortly before dementia onset can be a prodrome of dementia [43]. We classified individuals who were newly diagnosed with depression during the lagged-time period into the ‘no depression’ group. Second, we repeated the analysis using ICD-10 code disease definitions only. As the strict operational definitions of depression, CVD, and dementia can lead to selection bias, we applied mitigated definitions regardless of medication prescriptions or the number of hospital visits. All statistical analyses were performed using SAS 9.4 (SAS Institute Inc., Cary, NC/USA).