The result of the troponin I assay is typically evaluated in relation to a diagnostic cutoff. Clinicians might be reassured by results reported as negative, the diagnosis of myocardial infarction hinges primarily upon a positive result, and prognostic models that incorporate troponin most commonly use the test in a dichotomous way. Yet controversy remains (1- 2) regarding the proper placement of the cutoff. Some advocate that it be set at the 99th percentile (3) of a reference population, whereas others suggest that it should also be at a level where the coefficient of variation (CV) of the assay is <10% (4- 6), a higher level for most current assays. Whereas an increasing body of evidence suggests that any detectable troponin, even at levels below diagnostic cutoffs, might be associated with adverse prognosis (7- 12), an understanding of how “non-diagnostic” results should be interpreted in the acute setting, particularly when the possibility of myocardial infarction is not clear-cut, remains incomplete. Moreover, there has been little description to date of a quantitative relationship between troponin result and prognosis.

We sought to clarify the meaning of non-diagnostic results by studying troponin I as a predictor of all-cause inpatient mortality. In so doing, we developed a framework for considering troponin as a continuous predictor rather than as a dichotomous test.

We studied troponin I as a predictor of all-cause inpatient mortality, with retrospectively gathered electronic data from 2 clinically independent hospital centers. Subjects included every admitted patient who had troponin I sent from the emergency departments (EDs) of the 2 centers during the period from November 8, 2002 to January 31, 2005. Patients admitted to the psychiatric, pediatric, trauma, and hospice services were excluded. Days (or fractional days) with missing lab data due to known problems with data archiving were also excluded. The first troponin I sent from the emergency was pre-defined as the primary predictor. We also examined the highest troponin within 24 h and death within 24 and 48 h of arrival as alternate predictor and outcome variables.

We performed our study in 2 stages, deriving our model with data from one site and validating it with data from the other. All matching algorithms, hypothesis generation, and statistical methods were developed before gathering the validation data.

The 2 hospital centers are members of a common health care network and have a unified laboratory leadership, but the clinical departments are independent, and they have different patient demographics and different medical school affiliations.

We created the datasets by identifying all troponin tests performed in the hospitals and merging those records with data from our ED registration and financial systems (Eagle 2000, Siemens Medical Solutions Malvern, Pennsylvania), patient tracking systems (EmSTAT, A4 Health Systems, Cary, North Carolina), and electrocardiogram (ECG) retrieval system (MUSE CV, GE Marquette, Waukesha, Wisconsin). All data elements had previously been used for both quality improvement and research, and their accuracy has been verified (13). Troponin was assayed in heparinized plasma samples. The assays (Ortho-Clinical Diagnostics, Vitros Eci, Rochester, New York) were identical at the 2 hospital centers. Manufacturer-reported lower limit of detection was 0.038 μg/l, and cutoffs at the 99th percentile and 10% CV were 0.08 and 0.20 μg/l, respectively. The laboratory reported reference ranges as follows: “Negative” 0 to 0.08 μg/l; “Indeterminate” 0.09 to 0.20 μg/l; and “Positive” >0.20 μg/l. It reported results to a precision of .01 μg/l, even below the stated lower limit of detection. We initially studied mortality as a function of the negative, indeterminate, and positive categories, and noting a consistent difference in mortality between patients with a troponin of 0 and all other negatives, we considered 0 troponins separately.

Logarithmic transformation of troponin as a predictor in logistic regression analysis was first suggested by visual inspection of the smoothed mortality rates, plotted against troponin results. We then used the method of fractional polynomials (14- 15) to see whether another transformation might offer an improved fit. This is an automated process whereby a series of fractional power transformations are performed and compared in a nested stepwise fashion. This process consistently chose the log transformation over the untransformed variable and other power transformations.

Approximately one-half of our subjects had a measured troponin result of 0; this presented a problem with regard to the log transformation. Because a measured value of 0 is a censored value that represents an unknown distribution of concentrations between 0 and the lowest reported value of 0.01 and because the presence of circulating troponin might reflect a qualitatively different condition than situations wherein none is present, the placement of a 0 result on a continuous scale is somewhat arbitrary. We employed 2 methods for modeling the zeroes. For the results reported herein, we began by taking the non-zero troponin values (with the derivation set only), fitting a coefficient for log₁₀-troponin with only positive values, and then together with the known mortality rate for the 0 troponin population we derived the single best value to represent this distribution and replaced the zeroes with that value (0.004 μg/l). We used the same value to represent the zeroes in the validation set. We also tested a method that used a dummy variable to, in effect, fit 2 simultaneous models—a constant for troponin measured at 0 and log₁₀-troponin for positive values (15- 16). These methods yielded very similar results.

To assess the independence of troponin as a predictor of mortality, we identified other potential mortality predictors and then tested those that proved to be significant univariate predictors (p < 0.05 by Wald test) together with troponin in bivariate and then nested multivariate, logistic regression models. Independence of log₁₀-troponin was determined by stability of its odds ratio (OR) and by evaluation of goodness-of-fit in the bivariate and multivariable models. In cases where interaction was suggested by fluctuating ORs or residuals, we evaluated the fit of log₁₀-troponin on subsets of the data stratified by the interacting variable. Because they do not have an intuitive meaning in non-linear models (17), we do not report ORs for interaction terms. Putative mortality predictors that were tested included age, gender, creatinine, calculated glomerular filtration rate, arrival mode, chief complaint (of chest pain or shortness of breath), vital signs at triage (blood pressure, heart rate [HR], temperature, respiratory rate), and admitting diagnosis. Age was used as a continuous variable; other continuous variables such as creatinine, temperature, HR, respiratory rate, and blood pressure were dichotomized or converted to categorical variables on the basis of receiver operator characteristics curve analysis for optimal cutoff points (e.g., creatinine ≥1.3 mg/dl, HR ≥100 beats/min). Owing to differences in the dates of implementation of ED tracking systems at the different sites, vital sign data were available for only 78% of study subjects. A smaller fraction of other ECGs and creatinine values were unavailable (believed to be missing at random). In cases where nested models were studied, only subjects without missing data were considered.

To better understand the characteristics of the patient population and the accuracy of the admission diagnosis, we evaluated the International Classification of Diseases, 9th Revision (ICD-9)–coded discharge diagnoses by grouping the most frequently occurring diagnoses overall and then tabulating their frequency by admission diagnosis (rule out [r/o] acute coronary syndrome [ACS] vs. other). We also evaluated ICD-9 procedure codes (principal and secondary) to determine the fraction of patients who underwent diagnostic catheterization or coronary revascularization (percutaneous coronary intervention [PCI], coronary artery bypass graft), in the ACS and non-ACS admission diagnosis groups.

Data management and statistical analyses were performed with Stata 9.1 (StataCorp., College Station, Texas).

Of the 69,299 (26,704 derivation, 42,595 validation) patients admitted through the ED during the study period, 34,227 (12,135 and 22,092, respectively) had troponin measured, representing 48% (45% and 52%, respectively) of admitted patients overall. The distribution of patient characteristics across the 4 troponin categories and for non-study patients (patients admitted through the ED during the time period who did not have troponin sent before admission) is given in (Table 1).

Overall, 9,934 (29.0%) patients had an ACS admission diagnosis. Of these, 22.7% underwent diagnostic catheterization (vs. 3.5% for others and 0.61% for non-study patients) and 12.8% underwent a revascularization procedure (vs. 1.1% and 0.26%, respectively). The distribution of principal discharge diagnoses for patients admitted as possible ACS and for other study patients is given in (Table 2).

Among study patients, there were 1,291 deaths overall, giving mortality rates of 3.2% and 4.1% (derivation and validation, respectively). For non-study patients, mortality rates were 1.8% and 2.2%, respectively.

In-hospital mortality as a function of troponin result categorized as “zero,” “negative,” “indeterminate,” and “positive” were 1.8%, 3.3%, 5.0%, and 9.4%, respectively, for the derivation set and 2.3%, 3.9%, 6.6%, and 11.9%, respectively, for the validation set (Figure 1). For the combined data, the relative increase in odds of death compared with patients with troponins of 0 were: negative (but >0): 1.80 (95% CI 1.56 to 2.08); indeterminate: 3.04 (95% CI 2.53 to 3.65); positive: 5.81 (95% CI 4.89 to 6.91).

The unadjusted univariate OR for the prediction of death by log₁₀-troponin was 2.08 (95% CI 1.85 to 2.32) in the derivation set, 2.07 (95% CI 1.92 to 2.24) for the validation set, and 2.08 (95% CI 1.95 to 2.22) for the combined data (Hosmer-Lemeshow goodness-of-fit probability > chi-square[5] = 0.52, 0.16, 0.09, respectively). Thus, the odds of death approximately doubled for any 10-fold increase. The observed odds for a troponin of .01 compared with a troponin measured at 0 was 1.46 (95% CI 1.01 to 2.14) and 1.54 (95% CI 1.19 to 1.98) for the 2 data sets. (Figure 2)A and (Figure 2)B show the actual mortality rates versus troponin result (log-scale) superimposed upon the probabilities predicted by the univariate log-troponin model.

Of the putative mortality predictors tested, age, systolic blood pressure <100 mm Hg, creatinine ≥1.3 mg/dl, HR >100 beats/min, temperature ≥100.5 or ≤97°F, arrival by ambulance, white race, and complaint of shortness of breath were significant univariate predictors of increased mortality; normal ECG, chief complaint of chest pain, and Hispanic or black race were significant predictors of decreased mortality. Gender, HR <50 beats/min, and systolic blood pressure >150 mm Hg were not significant predictors. Calculated glomerular filtration rate performed in a manner similar to creatinine. The univariate ORs for each of the predictors together with the ORs for troponin in a bivariate model are given in (Table 3). The change in the OR for troponin with the addition of the “r/o ACS” term suggested the possibility of interaction—that the troponin OR is different for these patients than for others. Analysis of the combined data stratified by this term ratio yielded log₁₀-troponin ORs of 3.05 (95% CI 2.56 to 3.62) and 2.17 (95% CI 2.01 to 2.34) for ACS and non-ACS patients, respectively. The relationship between troponin and mortality in the ACS and non-ACS groups is illustrated in (Figure 2)C and (Figure 2)D. (Figure 3) demonstrates the observed versus predicted mortalities for troponin as a univariate mortality predictor among all study subjects, as stratified by admission diagnosis and together in a multivariable model with other significant predictors.

The area under the receiver-operating characteristic curve (AUC) was 0.78 for troponin alone in ACS patients versus .64 in others. McFadden’s pseudo R² statistic, which relates to the fraction of the outcome determined by the predictors in question, was 0.13 and 0.04. For multivariate models containing all predictors that remained significant in nesting, AUCs were 0.88 (ACS patients) and 0.78 (others), and pseudo R² were 0.14 and 0.28. Of note, these trends were similar for each of the other univariate predictors in the absence of troponin (e.g., age: AUC 0.76, R² 0.09 in ACS group; AUC 0.63, R² 0.02, others).

For 23,073 (66.8%) patients (66.8% derivation, 66.7% validation), a second troponin was sent within 24 h (median 540 min) of the first. Using the highest recorded troponin (within 24 h) as a predictor yielded ORs (2.00, 2.00) similar to those for the first.

Patients who did not have a second troponin (n = 11,350) had a higher overall mortality rate (4.36% vs. 2.62%, derivation; and 5.48% vs. 3.40%, validation) and were less likely to have an ACS admission diagnosis (13.8% vs. 38.2% and 8.6% vs. 38.2%, respectively). Log₁₀-highest troponin as a predictor of in-hospital mortality did not differ substantially from the first troponin, either when examined for the entire data or for the subset of patients who had a second troponin measured.

For patients who died, the median number of days between admission and death was 7 (interquartile range, 2 to 15). There were 210 deaths within 1 day and 331 deaths within 2. The ORs for death within 1 or 2 days as a function of log₁₀– troponin were 2.15 (95% CI 1.85 to 2.49) and 2.21 (95% CI 1.97 to 2.49), respectively, similar to that for any in-hospital mortality. Goodness-of-fit was also similar for these end points.

In our study populations, 48% of all patients admitted through the ED had troponin measured. This speaks to the degree to which our cohorts differ from those of most other published series (10- 11,18- 24), but the consistency between the 2 institutions suggests that our situation is not anomalous. Factors contributing to frequent troponin testing include the known existence of atypical ACS presentations, the preponderance of abnormal ECGs among admitted patients, the expedience of erring on the side of sending initial blood tests upon initial evaluation of ED patients, and the multiple and heterogeneously trained physicians caring for any given patient, particularly at teaching hospitals. Furthermore, the occasional finding of unexpectedly elevated troponin concentrations reinforces the tendency to measure them in atypical situations.

Our findings demonstrate that, in a heterogeneous population of acutely ill patients, the presence of troponin I, even at levels well below the 99th percentile of “normal,” is associated with an increased risk of mortality. In our population, there was a marked relative increase in mortality with the presence of any measurable troponin and there was no other discernible threshold for risk. The odds of mortality varied in direct proportion to order-of-magnitude increases, starting at the lowest detectible concentrations.

The mechanism for the troponin–mortality relationship is multi-factorial (12). It is clear that the relationship can exist in the absence of epicardial coronary disease (25- 26). Moreover, the extent to which the heart plays a mechanistic role in mortality and to which cardiomyocytes are “bystanders,” releasing troponin in response to systemic mediators, is unknown.

Previous studies have demonstrated variations in troponin distribution in apparently healthy populations, with age, gender, and ethnicity; the need for separate reference ranges has been proposed (6,27- 29). In our study, increased age and male gender were associated with a slight rightward skew in the distribution of troponin concentrations, but as in the well-characterized relationship between troponin, renal failure, and mortality (7- 8), the prognostic power of troponin was independent of these factors, thus arguing against separate reference ranges. This argument is also supported by the recent findings of Zethelius et al. (12).

Our study, along with other recent work, calls into question the very notion of a “normal reference range” for troponin I. Zethelius et al. (12) demonstrated a relationship between the presence of detectable troponin I levels below the 99th percentile of the overall reference population and the probability of long-term survival in a population of asymptomatic 70-year-old patients. The conundrum is that morbidity and death are “normal” events, and any reference population will contain a distribution of members with sub-clinical disease states. The ability of troponin I to discriminate between apparently normal individuals and the prognostic importance of levels well below the 99th percentile in the acute setting suggest that the use of a reference population for the purpose of establishing a diagnostic cutoff might be inappropriate.

To date, most outcome studies and clinical prediction instruments that incorporate troponin have used it in a dichotomous way. The loss of prognostic information that occurs when a continuous predictor is dichotomized by a single cutoff has been well described (30). To some extent, such use is driven by clinical usability. With the advent of comprehensive hospital information systems, however, it is likely that such risk assessment will increasingly be automated and the need for simplification might be diminished.

The logarithmic relationship that we outline has a number of important corollaries.

First, because the range of detectible values (0.01 to 0.08 μg) below the 99th percentile spans nearly an order of magnitude, a troponin result at the upper bound of what is called “negative” has a prognostic significance closer to that of most positive values than to 0. The model predicts, for example, an unadjusted mortality rate of 2.1% for a troponin of 0, a rate of 5.3% for a troponin of 0.08 μg/l (upper limit of 99th percentile, negative by all current recommendations), and 7.0% for a troponin of 0.21 μg/l (lower limit of positive by the strict 10% CV criterion).

Second, if 1 order of magnitude change in troponin concentration has the same prognostic implications as the next, then the untransformed variable cannot be used as a linear predictor and discussions regarding absolute changes in concentration will tend to be flawed. For example, the difference in mortality that we observed between patients with a troponin of 0.01 μg/l and those with 0 cannot be explained by a linear relationship but would be consistent with the logarithmic one if what is measured as 0 actually represents a distribution with a median of somewhere around 0.004 μg/l.

It is notable that currently recommended cutoffs are more stringent than the 97.5th percentile and 20% CV commonly used for other laboratory tests. This reflects conceptualization of troponin as a diagnostic gold standard, the attendant desire to avoid false positive tests, and the reality of the current assays’ imprecision at the low end. Our findings highlight the importance of improving assay precision. Whereas random precision errors at the low end cancel each other out in large study populations, such imprecision limits clinical utility at the level of individual patient care. For the assay under study, precision to the 0.01-μg/l level would be of clear benefit. Moreover, because the lower limit of what is truly “normal” has yet to be established, it is likely that an assay that could accurately detect even lower troponin levels would have additional value.

We have studied mortality as an outcome measure not only for its own importance but because it is unambiguous and because it obviates the problem of studying the specificity of a gold standard. Partly because diagnosis and workup are influenced by the test under study, we do not know with certainty which troponin elevations are associated with coronary disease but believe it likely that the relative odds of a “true acute coronary syndrome” would also relate logarithmically to troponin concentration. There are suggestions in previously published literature that the log-linear relationship might apply to asymptomatic populations, to other outcomes, and to troponin T as a predictor (12,18,22,24,31). The degree to which this holds true, however, awaits further study.

Although the increased OR for troponin in ACS patients compared with all others suggests that troponin release kinetics or the relationship between troponin and prognosis might differ in this population, it is also likely that the difference relates in part to an increased number of spurious deaths in the more heterogeneous non-ACS group (e.g., a death due to an idiosyncratic drug reaction or postoperative complication would not likely relate to an admission troponin) (32). This is supported by the finding that each of the other mortality predictors measured at the time of admission, including those that would be expected to be independent of coronary disease, were better predictors in the ACS group.

Because a second troponin was not measured in 34% of cases, we might underestimate the value of a peak troponin level as compared with the first. However, even when restricting the analysis to the subpopulation of patients who had 2 troponins measured, there was no significant difference between the first and the highest troponin as predictors. The mechanism for this lack of difference is likely multi-factorial, reflecting in part that the majority of subjects did not present with ST-segment elevation myocardial infarction, that a low first troponin implies early presentation and hence improved prognosis, that many ACS patients present hours after onset of symptoms, and that the prognostic value of troponin measured after primary PCI might be changed owing to a “washout” effect.

Although clinical troponin I assays measure a variety of different epitopes and vary widely in their scaling, the relationship between the results of different assays has been shown to be linear (33). Therefore, although our study employs a single assay, it would be expected that the log-troponin–mortality relationship would hold true for other assays, although the ORs might differ.

B-type natriuretic peptide has been described as a predictor of mortality in non-hospitalized patients without congestive heart failure (34). Because B-type natriuretic peptide was not measured in most of our subjects, we cannot know whether troponin’s prognostic utility would have been additive to B-type natriuretic peptides or whether it or some other factor not captured by us might diminish this utility. Although we could not identify any subset of patients whose OR for troponin was substantially diminished or whose fit of the logarithmic relationship was not good, the consistency of our findings in other sub-populations remains to be determined. Furthermore, although we did not find any non-zero risk threshold, it is possible that a future, more sensitive assay will identify some lower level of troponin that is not associated with increased risk.

Our findings support previous observations that the presence of any detectible troponin I is associated with an increased mortality risk and provide further evidence for the absence of any non-zero risk threshold. Mortality risk varies in direct proportion to order-of-magnitude changes in troponin I, more than doubling with any 10-fold increase. Troponin I (log-transformed) should therefore be viewed as a continuous prognostic indicator when assessing risk clinically and when building risk-assessment models. We suggest that current recommendations for diagnostic cutoffs and for the central role that any single cutoff assumes in the diagnosis of myocardial infarction merit reconsideration.

The authors thank Ms. Carmen Westen, Mr. Ernie Delia, and Ms. Mildred Jimenez for assistance with data collection.