Study design and participants

We analyzed prospectively collected data from the Healthy Aging Longitudinal Study in Taiwan (HALST). The details of the HALST were published previously [27]. In brief, the HALST is an ongoing prospective population-based cohort study in Taiwan that started in 2009 with 5663 community-dwelling adults aged 55 years and older enrolled. After being recruited, these participants received a baseline (first wave) examination and a follow-up (second wave) examination five years later. Sociodemographic status, lifestyles, diseases, and health status were assessed during these examinations. In addition, venous blood samples were collected, immediately processed, and maintained at -80 °C until assayed. This study was approved by the institutional review boards of the National Health Research Institutes and participating hospitals.

This study examined HALST participants without systemic inflammation, defined as a serum hs-CRP level of 2 mg per liter or higher [25, 28, 29], at baseline examination, leaving 3928 HALST participants entering this study. Supplementary Table 1 (of Additional file 1) showed the characteristics of the excluded participants. Each participant received measurement of leukocyte mtDNACN at baseline and measurements of serum hs-CRP at baseline and the follow-up examination.

Measurement of peripheral blood leukocyte mitochondrial DNA copy number

The leukocyte mtDNACN was measured in accordance with protocols described in previous studies [30]. In brief, we amplified a mitochondrial gene (ND1 gene) and a nuclear gene (β-globin gene) in the leukocyte total cellular DNA extracted from venous blood samples using a fluorescence-based quantitative polymerase chain reaction. The relative mtDNACN was estimated on the basis of these two genes’ threshold cycle numbers.

Measurement of serum high-sensitivity C-reactive protein

Baseline and follow-up serum hs-CRP levels were measured using a latex-enhanced immunoturbidimetric assay (ADVIA 1800 Chemistry system, Siemens AG, Munich, Germany). The intra-assay coefficient of variation was 5.61%, and the lowest detectable concentration was 0.12 mg/L.

Assessment of mortality

By linking to data from the Bureau of National Health Insurance of Taiwan death registry, any death events that occurred during the follow-up period were identified systematically, and the date of each death was ascertained. We followed each participant from the date of baseline examination (index date) until death or November 30, 2021, whichever occurred first.

Measurement of covariate variables

Other variables were measured during the baseline examinations. These variables were age (years), sex, education level (illiteracy, elementary school, junior high school, ≥ high school), smoking status (current, former, never), body mass index (BMI), dyslipidemia (serum triglycerides, serum high-density lipoprotein [HDL] cholesterol, serum low-density lipoprotein [LDL] cholesterol), and comorbidities (hypertension, cardiometabolic diseases [diabetes mellitus, stroke, cardiovascular disease], chronic kidney disease, lung disease). Further details are presented in the Supplementary Materials (Additional file 1: Supplementary Methods).

Statistical analyses

The baseline characteristics of the participants were summarized using descriptive statistics (mean [standard deviation] for normally distributed continuous variables, median [interquartile range] for nonnormally distributed continuous variables, and numbers [percentage] for categorical variables). Because both hs-CRP and the mtDNACN have a skewed distribution, both were analyzed as natural log-transformed values throughout the remainder of the analysis.

We jointly modeled the association of the leukocyte mtDNACN with longitudinal serum hs-CRP levels and time to all-cause mortality using maximum likelihood [31, 32]. The modeling details are presented in the Supplementary Materials (Additional file 1: Supplementary Methods). In the longitudinal part of the joint analysis, we modeled serum hs-CRP levels over time using linear mixed models (random-intercept-and-random-slope models). Each model incorporated fixed effects for the leukocyte mtDNACN, time (in years) since baseline examination, time’s interactions with the leukocyte mtDNACN, and random effects for intercept and slope with an unstructured variance–covariance structure. Departure from linearity was detected using restricted cubic splines with three knots. The correlations of longitudinal data for each individual were accommodated by the random effects in these models. We examined the directions and magnitudes of associations of the baseline leukocyte mtDNACN with hs-CRP based on parameter estimates.

In the survival part of the joint analysis, we modeled the association of the instantaneous hs-CRP level at time t (current value parameterization), as estimated by the longitudinal sub-models, with the risk of death at that time by using a Weibull parametric survival model. Based on the parameter estimates in the model, the hazard ratio (HR) per 1 standard deviation (SD) increase in log-transformed hs-CRP was estimated and presented.

Adjustments were made for covariates in the longitudinal (fixed effects for covariates and their interactions with time) and survival sub-models of joint analysis. As suggested in the literature, we identified a priori covariates that could act as potential confounders and a minimally sufficient set of covariates for adjustment by using a directed acyclic graph (DAG; Fig. 1) [33, 34]. In the longitudinal sub-model, the set of covariates was age, sex, educational level, smoking status, obesity (BMI), dyslipidemia, hypertension, and chronic kidney disease. In the survival sub-model, the set of covariates was age, sex, educational level, smoking status, obesity (BMI), dyslipidemia, hypertension, cardiometabolic diseases, lung disease, and chronic kidney disease. The joint models’ assumptions were examined using residual plots. The proportional hazards assumption of the survival part was evaluated using log–log survival curves and observed-versus-predicted survival curves.

Fig. 1

Directed acyclic graph depicting the assumed relationship among leukocyte mitochondrial DNA copy number, serum high-sensitive C-reactive protein, risk of death, and the potential confounders. Panel A displayed the relationships among leukocyte mitochondrial DNA copy number (exposure, represented by the green circle), serum high-sensitive C-reactive protein (outcome, represented by the blue circle), and covariates associated with both the leukocyte mitochondrial DNA copy number and serum high-sensitive C-reactive protein (ancestors of both the exposure and outcome, represented by the red circles). Panel B displayed the relationships among serum high-sensitive C-reactive protein (exposure, represented by the green circle), risk of death (outcome, represented by the blue circle), and the covariates associated with both serum high-sensitive C-reactive protein and risk of death (ancestors of both the exposure and outcome, represented by the red circles). The arrow from variable X to variable Y represents a direct effect of X on Y (an effect not mediated by other observed variables in the graph) that may have been present. The hypothesized causal pathway is presented in green. The biasing pathway is presented in red [34]

We conducted several sensitivity analyses. First, we added the baseline value of hs-CRP in the survival part of the joint analysis. Second, in the survival part of the joint analysis, we assumed that the risk of death at time t depends on the instantaneous hs-CRP level at that time. However, the risk of death at time t may depend on the levels of hs-CRP prior to time t. To examine whether the assumption might affect the results, we additionally modeled the association of the area under the hs-CRP trajectory curve (in mg/L × years) up to time t (cumulative effects parameterization), which captured the effects of the whole inflammation exposure history up to time t, with the risk of death at time t. Third, to minimize the effects of infectious processes, we excluded all participants with serum hs-CRP above 10 mg/L during follow-up (n = 82) and then repeated the analysis [35]. We then repeated the analysis again after excluding those with serum hs-CRP above 3 mg/L during follow-up (n = 319). Fourth, we conducted a sensitivity analysis to evaluate the robustness of our observations against potential unmeasured confounders by calculating the E-value [36]. E-value quantifies the minimal strength (on the risk-ratio scale) of the association of unmeasured confounders with both exposure and outcomes, conditional on the measured confounders, that can fully explain the observed exposure-outcome association.

Statistical significance was defined as a two-sided P value lower than 0.05. Analyses were performed using STATA, version 16.1 (StataCorp LP., College Station, TX, USA) and R, versions 4.1.2 (R Foundation for Statistical Computing).