Pre-transplant Dementia is Associated with Poor Survival After Hematopoietic Stem Cell Transplantation: A Nationwide Cohort Study with Propensity Score Matched Control (original) (raw)
Data Source
This study used data from Korea National Health Insurance Sharing Service (KNHISS, https://nhiss.nhis.or.kr/bd/ab/bdaba000eng.do). The Korean National Health Insurance Service (KNHIS), a compulsory public health insurance system of South Korea (hereafter “Korea”), provides universal coverage to all residents of Korea [21]. The KNHIS operates and manages KNHISS, in a form of National Health Information Database, which consists of healthcare data including health screening results, sociodemographic variables, and mortality for the whole Korean population. The data also contain all medical claims for the population covered under the KNHIS. Thus, almost all HSCTs conducted in Korea are included in this database. The database has been widely used in various epidemiological studies and is described in detail elsewhere [22,23].
Study Cohort
Patient selection for this study is summarized in Figure 1. We first identified a total of 8,230 patients who underwent HSCT between 2002 and 2018 using KNHIS database procedure codes X5061 for bone marrow and X5063 for peripheral blood HSCT. We first excluded 5,533 patients who were younger than 50 years. The remaining 2,697 patients were divided into those who were and were not diagnosed with dementia before HSCT (dementia group: n = 31; no dementia: n = 2,666). Thereafter, among 2,666 patients without dementia, 93 patients were selected as a matched control group (non-dementia group).
Outcome Variable and Covariates
The primary endpoint of this study was overall survival (OS) difference after HSCT according to pre-transplant dementia history. In order to utilize more conservative definition of dementia, patients who visited hospital due to International Classification of Diseases (ICD)-10 code of dementia (F00x, F01x, F02x, F03x, G30x, F051x, or G311x) were defined as having pre-transplant dementia. Demo-graphic and clinical characteristics based on HSCT-specific comorbidity index [24], which are available at KNHIS database, such as age, sex, baseline hematologic disease, previous non-hematologic or solid tumor, hypertension, diabetes, dyslipidemia, chronic obstructive pulmonary disease, cerebro- or cardiovascular disease, smoking, and social income were included as covariates.
Propensity Matching
In order to account for non-random treatment allocation, we used propensity score to select matched patients among 2,666 patients who were not diagnosed with dementia before HSCT. Propensity score adjustment enables the researcher to account for comparability between groups by balancing the distribution of biases and confounders [25]. Thus, it is one of the strongest methods to balance measured covariates for groups before analysis [26].
We first collected baseline characteristics of all 31 patients from the dementia group based on HSCT-specific comorbidity index (Supplementary Table 1; available online) [24]. Thereafter, the propensity was estimated for each participant in the dementia group using a multivariate unconditional logistic regression model. Next, using nearest neighbor matching, each patient in the dementia group were paired with a patient from the non-dementia cohort with the closest propensity score [27]. Age, sex, baseline hematologic disease, previous non-hematologic or solid tumor, hypertension, diabetes, dyslipidemia, chronic obstructive pulmonary disease, cerebro- or cardiovascular disease, smoking, and social income were all included as covariates, and matching ratio was 1:3 for each dementia-non-dementia group pair yielded total of 93 patients in the non-dementia group. In terms of caliper width, which is allowable amount of deviation in scores among matches, we set it as 0 of the logit of the propensity score, which was possible because of large cohort size [28].
Statistical Analysis
Difference between the two groups (dementia group and non-dementia group) in baseline demographic and clinical characteristics were compared using Student’s _t_test for continuous and chi-squared test for categorical variables. Retrospective cohorts were followed from the day they received HSCT to the occurrence of death or the last follow-up day (December 31, 2018), whichever came first. The OS rate represents the proportion of patients who were alive at a certain time after the date of transplantation and was associated with death due to any cause. OS rates were calculated using the Kaplan-Meier method and compared using the log-rank test. Variables with p < 0.01 in univariate analyses were entered into multivariate models with an exception for baseline age (50−59 years, 60−69 years, and above 70 years old). Finally, variables with p < 0.1 and baseline age (regardless of its p value) were included in multivariate Cox models of OS, using a backward stepwise model selection. For all statistical analysis, we used R statistical software (ver. 3.6.1; R Foundation for Statistical Computing, Vienna, Austria; 2019-07-05).