The ACC/AHA guidelines recommend accurate risk stratification of chest pain patients for initiating evidence-based therapies in the management of acute coronary syndromes (ACS) [1]. Risk stratification is achieved by using the various published risk scores and by clinical assessment. Troponin biomarkers represent an important part of the clinical assessment. They are widely used and are powerful prognostic predictors of adverse events in ACS syndromes. However, current risk stratification models like the Thrombolysis in Myocardial Infarction (TIMI) score, which though is simple to use may underestimate the predictive ability of powerful predictors like troponin biomarkers [2-4]. Accurate risk stratification is particularly important when stronger risk factor variables like troponin are present, but weaker ones are absent to give a low or intermediate score.

Another example of a risk prediction model in ACS is the Rush score [5-7]. It has a sound pathophysiologic basis that originates from Braunwald’s unstable angina classification [5-7]. The Rush score, developed from registry data, is a regression-weight based, validated instrument predicting mortality in ACS patients, but does not use troponin status [5-7]. In contrast, troponin biomarkers form an important component of most risk stratification models like the TIMI [8] and Global Registry of Acute Coronary Events (GRACE) [9] risk models.

We sought to increase the predictive performance of the Rush risk score, by first exploring the relationship between the Rush risk factors and troponin biomarkers and then incorporating troponin into the Rush risk factors as an additional component of this score. The rationale for selecting the Rush score for modification was that it is consists of easily ascertainable risk factors, is simple to use and it is one of the few scores that still do not use troponin. We use The Registry of Acute Coronary Syndromes (TRACS registry) for model development and validation.

Baseline characteristics of the TRACS registry population across the derivation and validation cohorts is shown in Table 3. The population mostly comprised of elderly, white, male patients with a high prevalence of diabetes, known CAD and hyperlipidemia, who were high risk, mostly non-ST segment myocardial infarction (NSTEMI) and unstable angina patients. There were no major differences between the randomly split derivation and validation cohorts (Table 3).

Of the total 2,866 patients, troponin biomarkers were positive in 1,339 (46.7%) patients, who were classified as having a NSTEMI. Of these NSTEMI patients, 950 (33.2%) patients had a first positive troponin on arrival in the emergency department, while the remainder 389 (13.6%) patients initially were troponin negative, but subsequently on repeat testing 6 hours and 12 hours apart, turned positive. The troponin negative patients (n= 1,527, 53.3%) were classified as having unstable angina. Of the 2,866 patients, the primary end-point of hospital mortality occurred in 75 (2.62 %) patients. The mean number of Rush risk factors were 2.24 ± 0.89 in the entire sample. Table 2 displays the in-hospital mortality by troponin status and sum of Rush risk factors in the derivation cohort.

Fig. (1) shows the distribution of the TRACS score and deaths (%) in each score sub-groups in the validation sample. A steep increase in mortality was seen with increase in TRACS score category. The TRACS score showed a good discriminatory ability for predicting in-hospital mortality.

ROC analysis demonstrated that area under the curve (AUC) was 0.7123 (95% CI 0.62-0.79) (Fig. 2). This was a significant improvement over the ROC AUC of the sum of Rush risk factors alone (AUC 0.6517, 95% CI 0.60-0.70). ROC analyses also showed a high predictive ability of the TRACS score in the sub-groups of diabetes, age ≥65, those with CAD, and troponin positivity (Table 4). We also calculated the TIMI score as described previously [8]. When ROC analyses were repeated for the predictive ability of the TRACS score across the traditional categories of TIMI score: low risk (TIMI score 0-3), intermediate risk (TIMI score 4-5) and high risk (TIMI score 6-7), the TRACS score had good discriminatory ability across all three traditional TIMI risk groups. It was reasonably good in the intermediate risk category of the TIMI score; an area where the clinician faces the most difficulty in risk stratification, and use of a separate score here shows statistical evidence of good predictive performance.

By univariate logistic regression analyses, the TRACS score categories were strongly predictive of death (OR per incremental category (0 to 5) was 1.24, 95%CI 1.14 – 1.34, p value <0.001). Hosmer and Lemeshow goodness-of-fit χ² was 2.53 and a p-value of 0.47, for five distinct quantile groups of the estimated probability of hospital mortality. This indicated a sound model fit. This model calibration is shown in Fig. (3). The TRACS score demonstrated little variation between the observed and expected prediction of death across its five score categories.

By multivariate analyses, the TRACS score independently predicted mortality (OR 1.14 per incremental category, 95% CI 1.01-1.28) even when adjusted for other known and well-documented, life-saving therapy covariates like revascularization, aspirin, beta blockers, heparin, and glycoprotein IIb/IIIa inhibitor use pre-cath and in the cathlab (Table 5). The Rush score risk factor sum was an important covariate of this model, and the TRACS score independently predicted mortality over and beyond that predicted by the RRF sum. Likelihood-ratio tests (LRT) between pairs of maximum likelihood logistic models were also performed to see if TRACS score had independent prognostic value. LRT was used to compare hierarchically nested models: the above full model shown in Table 5 and then a second nested model that dropped the TRACS score. This comparison resulted in a LRT χ² value = 4.91, p = 0.0267, indicating that the TRACS score was a powerful and independent predictor of in-hospital death. Since the full model included RRF sum as one of the model covariates, the above LRT result also implies that troponin status incorporated into the TRACS score with a weight of 3, contributes significantly, over and beyond the RRF sum, in the prediction of hospital mortality. A Bayesian approach to model comparison employing the use of the Bayesian Information Criterion (BIC) difference test revealed even better results. The BIC difference uses the Bayesian Information Criterion to estimate the probability that the model would be more likely after dropping one of the explanatory variables. Using the full model in Table 5, when each explanatory variable was dropped from the full model and BIC differences between the reduced (nested) model and the full (original) models were estimated for each of the explanatory variables in the full model, only the TRACS score had the largest BIC difference of 7.3, and BIC prime of 9.4 (p= 0.026). A BIC difference of 6-10 is considered “strong evidence” for variable selection according to Raftery [16]. Thus, troponin status with a weight of 3 incorporated in the TRACS score conveyed the highest information content in the prediction of hospital mortality, when directly compared with any other factor in the model.

Finally, Kaplan Meier survival analysis demonstrated that increasing TRACS score was associated with an increasing rate of mortality as a function of hospital stay (time). Fig. (4) shows that the Kaplan Meier curves for different score categories separated early and remained separate (log rank p<0.0001). Cox proportional hazards regression yielded a hazard ratio of 1.2 (p< 0.001, 95% CI 1.1 – 1.3) per incremental score category of the TRACS score.

These analyses suggest that using troponin status in a multiplicative manner with a weight of 3 adds incremental value in risk stratification for the prediction of hospital mortality, when used to improve the Rush score.

Similar exploratory analyses using the TIMI score [8] were also performed, although this was not the primary objective of this paper. When the TIMI risk score was recalculated using troponin in a multiplicative function with a weight of 4 (as determined via logistic regression model) this modified TIMI risk score had an ROC AUC of 0.6550 that was also a significant improvement over the original TIMI risk score (ROC AUC 0.5452, p value for difference = 0.0003).

Our data suggest that troponin status bears a multiplicative relationship to in-hospital mortality in risk stratification of ACS patients. We derive and validate this relationship to be a multiplicative factor of 3. The resulting TRACS score using this relationship can better risk-stratify an unselected, ED patient population presenting with chest pain, independent of that provided by the Rush score.

Using troponin status in such a multiplicative manner creates an exponential model of risk (Fig. 1) that closely approximates risk in real-world scenarios, where truly independent risk factors acting together multiply, rather than add risk. This creates a better separation of the risk profile in the intermediate zone, and consequently improved risk prediction. A potential downfall of such a strategy is to create widely differing and unstable estimates of risk across categories of the score. The Hosmer-Lemeshow χ² statistic, a measure of model stability and goodness-of-fit, does not demonstrate this. We hypothesize that this exponential risk modeling is a major reason for better risk prediction. In contrast, risk scores that assign a weight of 1 to troponin elevation, lose a major part of the prognostic information contained in troponin and is a reductive approach to the superior prognostic value of troponin.

Multiple large cohort and randomized studies also demonstrate that troponin is extremely important for prognostication. Recently a meta-analysis of 7 randomized trials and 11 cohort studies involving more than 10,000 patients, found that serum troponin had a three- to eightfold higher short-term risk of death for ACS patients [17]. Our data, showing a 3-fold risk is consistent with the meta-analysis data [17]. Thus, a greater emphasis on the information contained in troponin elevation for clinical decision making is a highlight of the TRACS score.

Other potential reasons account for the increased predictive ability of the TRACS score:

The TRACS score is a registry derived score, in comparison to other scores like the TIMI score that are derived and validated in clinical trial data. Non-participants in trials have a higher risk of adverse outcomes and greater heterogeneity than participants. Hence, it has been suggested that scoring systems be derived from observational data [4,18]. Our analysis from TRACS registry supports this observation. The strong discriminatory performance and superior calibration of the TRACS score indicate that it is likely to be useful in a “real-world” ACS population. Other studies done in the community reflect these observations. A community study demonstrated that AUC of the TIMI score was low for NSTEMI patients (c-statistic 0.59) [4]. Another community-based study comparing the TIMI, PURSUIT and GRACE scores identified that the TIMI score presented the lowest accuracy of the three risk scores (AUC 0.585) [2]. In this same study the GRACE score, derived from registry data, presented the best discriminatory accuracy, which was offset by its complexity that required the use of a computer program to calculate it. A Canadian registry of ACS patients [19] demonstrated that the GRACE score, derived from a less selected population compared to the PURSUIT score, provided superior calibration. These authors underscored the importance of risk model validation in the general ACS population rather than a clinical trial population to establish its generalizability before integration into clinical practice. This is consistent with our findings also. A meta-analysis of trials and cohort studies looking at the prognostic value of troponin, found that the magnitude of risk associated with a positive troponin was greater in cohort studies of consecutive patients than in clinical trials [17]. This again emphasizes that highly selected trial patients differ from unselected, heterogeneous, consecutive patients from cohort studies who have a greater potential for variation in underlying prognosis.