Risk prediction of atrial fibrillation and its complications in the community using hs troponin I

2.1 Study population

Analyses were based on data of four European, population-based, prospective cohort studies: FINRISK, Moli-sani, MONICA Northern Sweden, and Scottish Heart Health Extended Cohort (SHHEC). All study cohorts are part of the Biomarker for Cardiovascular Risk Assessment in Europe (BiomarCaRE) consortium and were approved by the local ethics committees. Informed, written consent was obtained from all participants. A more detailed description of the studies has been published previously.¹⁷

In total, our study population comprised 45,298 individuals, of which 13.2% were from the FINRISK cohort, 44.9% from Moli-sani, 18.6% from MONICA Northern Sweden, and 21.9% from SHHEC. Individuals with self-reported and physician-diagnosed history of AF as well as AF on the baseline ECG were defined as having prevalent AF and excluded from all analyses. Further, individuals with self-reported and physician-diagnosed history of heart failure, stroke, and cardiovascular disease including prior myocardial infarction were not included. Exclusion criteria and respective numbers are displayed in Table S1.

2.2 Data collection and biomarker measurements

At the time of inclusion in the studies, all participants completed a questionnaire, underwent physical examination, and provided a blood sample. The following risk factor information was collected: age, sex, body mass index (BMI), systolic and diastolic blood pressure, antihypertensive medication, smoking status, average alcohol consumption per day, diabetes status, serum-lipid concentrations with different fasting duration across cohorts, as well as pre-existing heart failure, coronary heart disease and stroke. Most variables were based on history. BMI and blood pressure were measured during the examination. During follow-up, interim disease onset and outcomes were identified from national hospital discharge register partially including outpatient visits, national causes of death registers and the national drug reimbursement registers. The diagnosis of incident AF was based on study ECG tracings, questionnaire information, national hospital discharge registry data, ambulatory visits to specialized hospitals, and screening of death registry data. Regarding the Moli-sani study, the assessment of events was made by record linkage with hospital discharge registers and national causes of death registers. The outcomes were AF, heart failure, stroke, cardiovascular events (including first fatal or non-fatal myocardial infarction or stroke event) and total mortality. These data were harmonized for joint analysis in the MORGAM Project.¹⁸

Blood samples were stored at −80°C or in liquid nitrogen at −196°C (Moli-sani) and freeze–thaw cycles were minimized. hsTnI concentrations were measured centrally in the BiomarCaRE core laboratory using a hsTnI immunoassay (Abbott Diagnostics, ARCHITECT i2000SR). NT-proBNP levels were measured on the ELECSYS 2010 and the Cobas e411 using an electrochemiluminescence immuno-assay (ECLIA, Roche Diagnostics) and high-sensitive C-reactive protein (hsCRP) was determined by latex enhanced immunoassay (Abbott, Architect c8000).

2.3 Statistical methods

Continuous variables were presented as median (25th, 75th percentile) and binary variables as absolute and relative frequencies. To study the associations of AF risk factors with time to AF, multivariable Cox regressions were performed. First, univariable Cox regression was performed to study the association between AF and of high sensitivity troponin I (hsTnI) quarter. Second, our primary multivariable model (Model I) included the following covariates not considered to be time-varying and measured at baseline: sex, BMI, log10(serum triglycerides), alcohol consumption, systolic blood pressure, total cholesterol minus high density lipoprotein (HDL) cholesterol, current smoking, prevalent diabetes, and MORGAM geographic area (see Table S2) and with age as a time-varying covariate (the analysis dataset was segmented into yearly time-intervals and age was updated yearly using the start value it had at each interval). Time since baseline was used as the time scale in these Cox regressions. We selected the strongest biomarker from Cox regressions for further assessment of predictive ability in comparison with hsTnI. 12,036 of the 45,298 hsTnI values were below the assay range but were included in our analysis. Values of 0.0 ng/L were replaced by a small positive value (0.025 ng/L) in order to be able to perform log transformation. When hsTnI was included in a Cox regression, it was included as a continuous variable.

The primary model (Model I) was extended as follows: +log10(NT-proBNP) (Model II), +log10(hsTnI) (Model III), and +log10(NT-pro-BNP) + log10(hsTnI) (Model IV). In Models I to IV, the linearity of the model was assessed by testing all possible second-order interactions between covariates and adding them to the models I to IV if they reached significance in one or more models, taking into account multiple testing using Bonferroni correction. The proportional hazards assumptions were tested by looking at the correlation between transformed survival time and scaled Schoenfeld residuals as implemented by using the R function cox.zph()¹⁹ and interactions with time were added to the model if needed to make the proportional hazards assumption valid. If it was needed to add an interaction to one or more of the Models I to IV, it was automatically also added to the other models of Models I to IV. Continuous variables were always centred at their mean when included in an interaction. The ultimately obtained Models I to IV were tested with cox.zph() to be free of PH-violations and also the Hosmer-Lemeshow test showed no significant problems with the goodness-of-fits of these models. In supplemental analyses, we also fitted multivariable models without interactions. We performed multivariable models where only age, sex, and MORGAM geographic area remained as adjustments.

In secondary analyses, Cox regressions were performed for the hazard ratio (HR) of log10(hsTnI) for overall mortality, incident heart failure, incident cardiovascular disease, and incident stroke with the same adjustment as in our primary model (Model I).

Cumulative incidence curves where generated. The p-value for the difference between cumulative incidence curves was calculated using the log rank test.

Harrel's C-statistic, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated for comparing the predictive performance of linear predictors calculated using the multivariable-adjusted beta coefficients obtained in this study. The predictive abilities for incident AF were analysed using four different models: the beta coefficients of CVRF alone; the beta coefficients of CVRF and log10(hsTnI); the beta coefficients of CVRF and log10(NT-proBNP); the beta coefficients of CVRF, log10(NT-proBNP) and log10(hsTnI). Confidence intervals and p-values for differences between C-statistics were estimated using bootstrapping with 10,000 replications.^{20, 21} Each individual was repeated 10 times in the bootstrap population. Samples were drawn from this bootstrap population using simple random sampling.

A two-sided p-value of <.05 was considered statistically significant. Analyses were performed with R version 3.5.3.²²

3.1 Baseline characteristics

The median age of the overall study cohort was 51.4 (lower and upper quartiles 43.1, 59.8) years, 45.0% (N = 20,375) were men. During a median follow-up time of 7.7 (lower and upper quartiles 4.2, 17.3) years N = 1734 (3.8%) of the 45,298 participants developed AF. Median time to their first AF event was 11.8 (lower and upper quartiles 4.9, 17.4) years. In total 4108 participants died during the follow-up time, which equals a mortality of 9.1%. Further baseline characteristics are shown in Table 1 as well as Tables S3 and S4.

3.2 AF prediction models

The study population was divided into four equally sized groups depending on hsTnI values. In univariable Cox regression, those in the highest quarter of hsTnI values at baseline (≥4.2 ng/L) had a 3.91-fold (95% confidence interval [CI] 3.30, 4.63; p < .01) higher hazard for developing AF compared to those in the lowest quarter (<1.4 ng/L). Cumulative incidence curves for AF by quarters of hsTnI and the respective numbers at risk are displayed in Figure 1.

Cox proportional hazard models adjusted for age and sex revealed a significant association between log10(hsTnI) and AF (HR per 1 SD [=0.50] increase in log10(hsTnI) 1.16; 95% CI 1.09, 1.23; p < .01) and in multivariable-adjusted Cox proportional hazard models also a statistically significant association was seen (HR per 1 SD [=0.50] increase in log10(hsTnI) 1.08; 95% CI 1.01, 1.16; p = .03). In multivariable-adjusted Cox proportional hazard models log10(NT-proBNP) (HR per 1 SD [=0.41] increase in log10(NT-proBNP) 2.11; 95% CI 1.98, 2.25; p < .01) as well as log10(hsCRP) (HR per 1 SD [=0.48] increase in log10(hsCRP) 1.09; 95% CI 1.03, 1.16; p < .01) were statistically significantly related to incident AF. Hazard ratios for hsTnI, NT-proBNP and hsCRP in different Cox proportional hazard models are presented in Figure 2 and Table S5.

For CVRF alone, the C-index was 0.81 with bootstrapped standard deviation of 0.01. For the model comprising CVRF and log10(hsTnI) improvement of risk prediction compared to CVRF alone was found. Addition of log10(NT-proBNP) as the strongest biomarker resulted in significantly improved model discrimination (C-index 0.84, ΔC 0.026; 95% CI 0.019, 0.032; p < .001). The combination of CVRF and both biomarkers did not show higher predictive ability compared to the model with CVRF and log10(NT-proBNP) alone. However, categorical NRI (with categories based on the quartiles of linear risk predictors) was highest (14.5%), when comparing the model with CVRF alone and the model with the combination of CVRF, log10(NT-proBNP) and log10(hsTnI). IDI was highest (0.6218), when comparing the model with CVRF alone and the model with the combination of CVRF, log10(NT-proBNP) and log10(hsTnI). C-indices, NRIs and IDIs are shown in Table 2 and for each geographical area separately in Table S6.

We further examined the predictive ability of log10(hsTnI) for overall mortality and the AF complications heart failure, cardiovascular events and stroke after newly diagnosed AF. Higher hsTnI concentrations were significantly associated with overall mortality and most prominent with heart failure (HR per 1 SD increase of log10(hsTnI) 1.37; 95% CI 1.12, 1.68; p < .01), as presented in Table 3.

4.1 Strengths and limitations

The four included population-based, prospective cohort studies are mainly composed of participants of European ancestry. Therefore, our findings may not be generalizable to other ethnicities. Likewise, the exclusion of individuals with previously diagnosed heart failure and myocardial infarction may limit generalizability. The outcomes were largely defined using national hospital-based and drug reimbursement-based databases. However, AF is often an outpatient diagnosis. Thus, we may have missed AF cases with outpatient diagnosis only. However, follow-up during the last years also included out-patient data and may have reduced this source of bias. A major strength of the study is the large number of individuals with harmonized baseline data across Europe in combination with a long follow-up.