Establishing the 99th percentile of a novel assay for high-sensitivity troponin I in a healthy blood donor population

Introduction

Suspected acute myocardial infarction (AMI) is a frequent cause for admission to the emergency department. To clarify whether an AMI is present one usually measures the patient’s cardiac troponin (cTn) concentration, and the cut-off for diagnosing AMI is the 99th percentile of cTn among healthy individuals [1]. To establish the 99th percentile in a healthy population, individuals using cardiac drugs or with underlying disease, elevated values of natriuretic peptide or an estimated glomerular filtration rate (eGFR) <60 mL/min/1.73 m² should not be included [2], [3]. Non-parametric methods for estimating percentiles should be used on values measured from at least 300 women and 300 men [2], [3], [4], [5]. Further, for high-sensitivity (hs) assays it is recommended to establish sex-specific 99th percentiles and report them with patient results in clinical practice [1], [2], [3]. Assays for cTn measurements are not standardized. Ideally, each laboratory should establish the 99th percentile relevant for its method and population. Despite these recommendations, there is no universal consensus regarding how healthy reference individuals should be recruited or how outliers should be detected.

Recently, Siemens Healthcare Diagnostics released a new assay for cTnI, the ADVIA Centaur^® High Sensitivity Troponin I assay (hs-cTnI). According to the manufacturer, the common 99th percentile in healthy men and women is 47 ng/L and somewhat lower, i.e. 40 ng/L, according to a recent Italian study [6]. They both used frozen samples and the populations were compound, consisting of both blood donors, assumingly healthy patients from primary health care and hospital employees. The aim of the present study was to determine the 99th percentile for the ADVIA Centaur^® High Sensitivity Troponin I assay in a prospective study using fresh plasma samples from a healthy reference population.

Materials and methods

Results

A total of 1000 blood donors were included, of which 497 were women (Table 1 and Figure 1). The median time from sample collection to analysis was 190 min, the 75th percentile was 308 min and range was 55 min to 29 h. The distribution of hs-cTnI is presented in Figure 2. Gender, age, eGFR, HbA_1c and NT-proBNP were not associated with the 99th percentile for hs-cTnI in the quantile regression analysis, neither in a multivariable model nor in univariable models (p>0.10). The association between age and hs-cTnI is illustrated in Figure 3.

Figure 1:

Age distributions.

Histograms illustrating age distributions among females and males.

Figure 2:

Distribution of hs-TnI.

Dotplots illustrating the distributions of hs-cTnI among females and males on a logarithmic scale.

Figure 3:

Hs-TnI and age.

Scatterplot of hs-cTnI on a logarithmic scale and age.

The 99th percentile (90% CI) for hs-cTnI was 96 (65–149) ng/L for the total population, 59 (33–107) ng/L for women and 144 (92–223) ng/L for men (Table 2). One donor had eGFR <60 mL/min/1.73 m², eight had HbA_1c >47 mmol/mol and eight had NT-proBNP above the upper reference limit. When excluding these 17 donors the 99th percentile was 95 (61–149) ng/L, which was not significantly different from the total population. By repeated Reed’s test without transformation, one observation was considered an outlier in the total population and in the distribution among women. Excluding this observation lowered the 99th percentile for women by 20 ng/L but did not change the 99th percentile for the total population (Table 2). The Gaussian distribution functions estimated on the basis of the 45% (n=447) of log-transformed cTnI values above the LoQ corresponded reasonably well with the actual results distributions above the LoQ (Supplementary Figure 2 and Supplementary Table 4). Thirteen, three and 12 observations were considered outliers according to Tukey’s test in the whole population, among females and among males, respectively. Excluding these observations had a significant impact on the estimated 99th percentiles (Table 2).

Hs-cTnT was measured in the same samples in order to validate the healthy status of the donor population (Figure 4). Sixty-five percent (n=655) of measured values were above the LoQ. The 99th percentile for hs-cTnT was 14.7 (13.6–19.2) ng/L for the whole population, 9.1 (7.8–15.1) ng/L for women and 16.8 (14.3–30.3) ng/L for men. Excluding the 17 participants with high HbA_1c, high NT-proBNP or low eGFR did not influence the 99th percentile significantly (Table 2). No outliers were identified by Reed’s test. Five, three and four observations were considered outliers by Tukey’s test on the basis of the estimated Gaussian distributions in the whole population, among women and among men, respectively. Excluding these observations lowered the 99th percentile moderately. The correlation between hs-cTnI and hs-cTnT was moderate (Supplementary Figure 1), although statistically significant (Spearman’s ρ=0.56, p<0.001).

Figure 4:

Distribution of hs-TnT.

Dotplots illustrating the distributions of hs-cTnT among females and males on a logarithmic scale.

Discussion

We estimated a 99th percentile for hs-cTnI of 96 ng/L for healthy blood donors, 59 ng/L for women and 144 ng/L for men. This is significantly higher than estimates reported by the manufacturer (Table 2), which are 47, 37 and 57 ng/L, respectively. Our estimates are more than twice as high as reported by the manufacturer for the whole population and for men. The manufacturer used Reed’s test to eliminate outliers. Corresponding estimates after removing outliers identified by Reed’s test were 93, 39 and 144 ng/L.

To the best of our knowlwdge, only one other publication of the 99th percentile exists for this hs-cTnI method [6]. In that study of an Italian cohort a 99th percentile for hs-cTnI of 40 ng/L was reported for the total population, 32 ng/L for women and 43 ng/L for men, which is more in accordance with the manufacturer. Tukey’s test for outliers was applied, apparently with the limit Q3+3×IQR, and not 1.5×IQR as recommended by others [4], [11], [12]. In addition, IQR was apparently estimated on basis of troponin values far below the LoQ, and even below the LoB. We do not know what impact this way of handling outliers had on the estimated 99th percentiles, as this is not reported.

As our troponin results were heavily skewed and contained a significant number of values below the LoQ, results could not be converted to fit an approximate Gaussian distribution, which is a requirement for using Tukey’s test [4], [12]. In an attempt to overcome this, the underlying Gaussian distributions were estimated and used to calculate Tukey’s limits for outlier detection, as explained in the Methods section. Excluding outliers outside these limits significantly lowered the estimates for the 99th percentiles to 53, 35 and 55 ng/L for the whole population, for women and for men, respectively. No outliers were identified when we used the limit 3×IQR, as in the Italian study [6]. Even though these estimates for the 99th percentiles are in the same order of magnitude as reported by the manufacturer and in the Italian study, there is a higher uncertainty associated with the limits for outlier detection estimated in this way, and this may influence the validity of the estimated 99th percentiles.

Although samples were analyzed as soon as possible after sample collection, six samples were not analyzed until the next day due to technical issues. Sample stability of hs-cTnI is good, as documented by the supplier. Results from these samples were therefore included in our estimates. Excluding them changed the estimates for the 99th percentiles of hs-cTnI less than 1 ng/L.

Based on the information available and the fact that the results distribution for hs-cTnT in our donors was not different from other published studies [14], we do not believe that any of the high results for hs-cTnI observed in our study are due to relevant and unrecognized comorbidity, incorrect sample handling or measurement errors. Hence, there are no good reasons to exclude observations on any other than pure statistical grounds. We believe that our results reflect the biological variation of cTn in a healthy blood donor population. But as presented in Table 2 and as demonstrated by others, removing outliers based on statistical considerations may have a significant impact on the estimates [15], [16].

One possible explanation for the differences observed for the 99th percentile in our study compared to the manufacturer’s and the Italian study [6] is differences between the populations. Our reference sample group consisted of 1000 apparently healthy Norwegian blood donors aged 18–75 years old, while the manufacturer included 2010 blood donors and apparently healthy patients from primary care practices in the US aged 22–91 years old. The Italian population comprised blood donors, apparently healthy individuals from the laboratory staff, clinical staff and participants in a screening study of subclinical cardiovascular disease. The age ranged from 18 to 86 years old (mean 51 years old). In all three studies approximately half of the participants were women. As our donors tended to be younger than in the other two studies, different age distributions are likely not the only explanation for observed discrepancies in the 99th percentiles for hs-cTnI. Due to younger age, the prevalence of unrecognized comorbidity could also be expected to be lower in our population. Four of the participants, having some of the highest hs-cTnI values, were offered to have cTn measured in a second sample and a cardiology check-up including electrocardiography and echocardiography. All donors were considered to be healthy and without cardiac disease. However, they all reported to have been working out before sample collection. It is well known that cTn can increase in relation to strenuous physical activity [17], [18]. As physical activity must be considered normal in a healthy population, and there are no recommendations to exclude donors based on this criterion when estimating the 99th percentile of cTn [2], [3], including donors that are more physically active than other populations can, almost as a paradox, cause higher estimates for the 99th percentile for cTn. In turn, this may influence the diagnostic properties of cTn by reducing the diagnostic sensitivity for AMI.

According to the APACE study [19] the ROC curves for AMI for the hs-cTnI assay from Siemens, the hs-cTnI assay from Abbott and the hs-cTnT test from Roche are very similar (figure 2 and Supplementary figure 3 in [19]). However, when using each of the assay’s 99th percentile as cut-off the sensitivity and specificity differs (table 4 in [19]), i.e. 70% and 93% for hs-cTnI from Siemens, 78% and 93% for hs-cTnI from Abbott and 94% and 78% for hs-cTnT from Roche, respectively. In other words, the generally accepted 99th percentiles represent different points on the ROC curve, which could be due to that the cut-offs have been established in different populations or that outliers have been handled differently. One could argue that the rise and/or fall in cTn as a criterion for AMI diminish these differences [1]. However, according to the APACE study, combining hs-cTnI at admission with results at later time points did not increase the diagnostic accuracy significantly (Supplementary table 3A in [19]).

A weakness of our study is the relatively low number of observations, although well above the recommended minimum of 600 participants [2], [3], [4], [5]. This also results in a low power to detect associations between hs-cTnI and potential predictors in quantile regression of the 99th percentile. Another weakness is that we did not examine all the participants with electrocardiography, echocardiography or other diagnostic imaging studies to uncover subclinical cardiac disease, as this was beyond our capabilities.

This is the first published study of the 99th percentile for this hs-cTnI assay using fresh samples. We performed the study in a pragmatic way, feasible for most clinical laboratories. Our population is well defined and can easily be reproduced by other laboratories. We also present results and the handling of outliers in a transparent way. As we also measured hs-cTnT, among other analytes, in the same samples, donors with relevant comorbidity could be identified. The hs-cTnT results were not different from those reported by the manufacturer and other relevant studies [13], [14], indicating that we studied a healthy population in line with previous reports. Further, removal of potential outliers did not affect the cut-off for hs-cTnT significantly.

Our results indicate that current recommendations for establishing the 99th percentile of cTn can lead to very different estimates. The consequence could be that different cut-offs for AMI are used, which might influence patient care significantly. A more detailed recommendation of how to define a healthy reference population in this setting should be established, including how to deal with outliers and the potential effects of physical activity. The question should also be raised whether the time has come to stop using the 99th percentile as the diagnostic cut-off for AMI, as it is notoriously difficult to estimate precisely in a consistent way, and instead rely on optimal cut-off values identified by receiver operating characteristic (ROC)-analyses in clinically relevant populations [20], [21].