Predicting risk of cardiovascular events 1 to 3 years post‐myocardial infarction using a global registry

Abstract Background Risk prediction tools are lacking for patients with stable disease some years after myocardial infarction (MI). Hypothesis A practical long‐term cardiovascular risk index can be developed. Methods The long‐Term rIsk, Clinical manaGement and healthcare Resource utilization of stable coronary artery dISease in post‐myocardial infarction patients prospective global registry enrolled patients 1 to 3 years post‐MI (369 centers; 25 countries), all with ≥1 risk factor (age ≥65 years, diabetes mellitus requiring medication, second prior MI, multivessel coronary artery disease, or chronic non‐end‐stage kidney disease [CKD]). Self‐reported health was assessed with EuroQoL‐5 dimensions. Multivariable Poisson regression models were used to determine key predictors of the primary composite outcome (MI, unstable angina with urgent revascularization [UA], stroke, or all‐cause death) over 2 years. Results The primary outcome occurred in 621 (6.9%) of 9027 eligible patients: death 295 (3.3%), MI 195 (2.2%), UA 103 (1.1%), and stroke 58 (0.6%). All events accrued linearly. In a multivariable model, 11 significant predictors of primary outcome (age ≥65 years, diabetes, second prior MI, CKD, history of major bleed, peripheral arterial disease, heart failure, cardiovascular hospitalization (prior 6 months), medical management (index MI), on diuretic, and poor self‐reported health) were identified and combined into a user‐friendly risk index. Compared with lowest‐risk patients, those in the top 16% had a rate ratio of 6.9 for the primary composite, and 18.7 for all‐cause death (overall c‐statistic; 0.686, and 0.768, respectively). External validation was performed using the Australian Cooperative National Registry of Acute Coronary Care, Guideline Adherence and Clinical Events registry (c‐statistic; 0.748, and 0.849, respectively). Conclusions In patients >1‐year post‐MI, recurrent cardiovascular events and deaths accrue linearly. A simple risk index can stratify patients, potentially helping to guide management.

Results: The primary outcome occurred in 621 (6.9%) of 9027 eligible patients: death 295 (3.3%), MI 195 (2.2%), UA 103 (1.1%), and stroke 58 (0.6%). All events accrued linearly. In a multivariable model, 11 significant predictors of primary outcome (age ≥65 years, diabetes, second prior MI, CKD, history of major bleed, peripheral arterial disease, heart failure, cardiovascular hospitalization (prior 6 months), medical management (index MI), on diuretic, and poor self-reported health) were identified and combined into a user-friendly risk index. Compared with lowest-risk patients, those in the top 16% had a rate ratio of 6.9 for the primary composite, and 18.7 for all-cause death (overall c-statistic; 0.686, and 0.768, respectively). External validation was performed using the Australian Cooperative National Registry of Acute Coronary Care, Guideline Adherence and Clinical Events registry (c-statistic; 0.748, and 0.849, respectively).  5 Given the recent availability of new therapies that can improve outcomes in these patients, [6][7][8] there is a need to understand the drivers of risk in this cohort so as to identify those at greatest absolute risk who are likely to sustain the greatest absolute benefit.
This article focuses on incidence of the composite primary endpoint (MI, unstable angina requiring revascularization, stroke, and all-cause death) over 2 years, and characterizes any influence by baseline patient factors. Our goal was to develop a user-friendly risk index incorporating readily available items, see how strongly it facilitated risk discrimination for the primary endpoint and for allcause death, and validate against an external population. Finally, we discuss how this risk index may be used in patients with stable coronary disease to determine those with relatively good prognosis and those at higher risk who may benefit from more intensive management.

| HYPOTHESIS
That a practical cardiovascular risk index could be developed using data from a global registry of patients >1 year post-MI, followed up for 2 years.

| Study design and patients
TIGRIS is a prospective, global registry of patients enrolled 1 to 3 years post-MI in 25 countries in Europe, North America, Latin America, Asia, and Australia, followed for 2 years. The study design and patient characteristics have been described. 5,9 Eligible patients had at least one of the following: age ≥65 years, diabetes mellitus requiring medication, a second prior MI, angiographic evidence of multivessel disease, and chronic non-end-stage renal dysfunction. Eligible patients who had survived an acute coronary syndrome (ACS) were enrolled at discharge from participating centers and subsequently contacted every 6 months by phone or at study sites to ascertain outcome events and changes in medications. All outcome events reported by patients and relatives (eg, hospitalizations, cardiovascular events, deaths) were confirmed by the study sites.

| Risk index development
The predefined primary composite endpoint of MI, unstable angina with urgent revascularization (UA), stroke, and all-cause death showed a linear accumulation of events over time. 10 Hence, incidence of the primary outcome by baseline variables was reported as rate per 100 patient-years. Poisson regression models simultaneously estimated the association of several baseline variables with risk of events expressed as incidence rate ratios (IRRs) and 95% confidence intervals.
We used forward stepwise variable selection to derive a preliminary multivariable predictive model for risk of the primary outcome.
The five high-risk eligibility criteria were forced into this model, as were sex and geographic region. All other variables needed to achieve P < .05 for inclusion. A final condensed Poisson model for the primary endpoint was obtained by only including variables significant at P < .01, and by modeling age (<65 and ≥65 years) and diabetes as binary variables.
To assess the risk impact of self-reported health, the 3-level EuroQoL-5 dimensions (EQ-5D-3L) survey instrument 11 was used.
Patients graded five dimensions (mobility, self-care, usual activities, pain/ discomfort, and anxiety/depression) as no, moderate, or severe problem scoring each as 0, 1, or 2 points, respectively, and summing to yield a simple overall score (range 0-10). Recommended EQ-5D scorings such as UK-weighted index 12 are complex and impractical for clinical use. The final predictive model was converted into an integer risk index: variables with rate ratios ranging from 1.33 (congestive heart failure) to 1.69 (prior major bleed) were each assigned 1 point, while the strongest predictor EQ-5D-3L overall score ≥4 with rate ratio 2.06 was assigned 2 points. This risk index was formed into six ordered categories from 0 points to ≥5 points. Risk discrimination was quantified using incidence rates, IRRs using 0 points as reference, and Kaplan-Meier plots over 2 years. Harrell's c-statistic summarized the strength of discrimination. 13

| External validation
External validation of the risk index used the Australian Cooperative National Registry of Acute Coronary Care, Guideline Adherence and Clinical Events (CONCORDANCE), which included 4672 post-MI patients seen 6 months post-discharge. 14 Occurrences of the primary composite outcome and death were documented at 6 months (n = 3197) or 18 months (n = 1451). Precise event dates were unavailable with logistic regression used to predict outcomes, adjusting for duration of follow-up. EQ-5D was missing in 1588 patients in CONCORDANCE. We therefore used multiple imputation using five imputed datasets, combining results using Rubin's rules 15 ; predictor variables were the primary outcome, death, and other risk model variables.
All statistical analyses used Stata version 15.1.

| Primary outcome
The primary outcome occurred in 621 (6.9%) from 9027 patients over For the primary outcome and components, a steady linear accumulation of events occurred over time. 10

| Identifying predictors of risk
Univariate associations of patient characteristics to the primary outcome incidence rate are provided in Table S1, with IRRs shown both unadjusted and adjusted for the five eligibility criteria, region, and country. Unadjusted incidence rates by region and country are shown in Table S2.
To identify which variables remained statistically significant independent risk predictors in multivariable analyses, forward stepwise variable selection was used to derive a preliminary predictive model for risk of the primary outcome ( Table 1). The influence of age is effectively summarized by elevated risk for age ≥65 years. Diabetes is an important risk predictor, particularly for insulin-treated patients, as are having a second prior MI, a prior major bleed, peripheral arterial disease, and prior heart failure. Also, cardiovascular hospitalization in the last 6 months, medical management only for the index MI, and diuretic therapy at enrollment also carried elevated risk.
The EQ-5D findings indicate the prognostic importance of patient-reported heath status with each constituent item showing increased univariate trends of risk from no-to-some-to-severe problems (Table S1). Their sum yielded an overall ED-5D-3L score ranging from 0 (no problems on all five items) to 10 (severe problems on all five items). An overall score of 3 has a significantly elevated risk, which increased further for a score of ≥4 (Table 1). The EQ-5D visual analog score did not independently predict risk.
Sex, multivessel disease, and region were not statistically significant independent predictors. Patients on single antiplatelet therapy at enrollment appeared to have a lower risk. The underlying selection processes are unknown so this factor was not considered further. Variables expected to predict risk (eg, low blood pressure, elevated heart rate) were not independent predictors.

| Final predictive model
A final refined predictive model for the primary outcome is shown in To make this prediction model more user-friendly, we propose a risk index for each patient ( Table 2). All items contributed 1 point to the risk index, except for EQ-5D overall score ≥4 (2 points). Distribution of the risk index for 8978 patients with complete information is shown in Figure S1. A value of 0 points occurred in 12% of patients, 1 point was most common (33% of patients), followed by a skewed distribution to a maximum of 10 points in four patients.  Figure S2). Figure 2A shows cumulative incidence of the primary outcome by categories of the risk index, revealing marked separation in risk. The 2-year cumulative incidence for the primary outcome ranged from 2.8% for 0 points up to 11.9% and 23.7% for 4 and ≥5 points, respectively. For all-cause death, compared to patients with 0 points, there is an even steeper risk gradient (IRR) ranging from 1.77 for 1 point to 11.19 and 27.61 for 4 and ≥5 points, respectively (Figures 1 and 2B). c-Statistics for the risk index are 0.686 for the primary outcome and 0.768 for allcause death.  (Table S3). When applying the TIGRIS risk index to CONCORDANCE, we observed similar, markedly steep gradients for both the primary composite outcome and all-cause death (Figure 3).   Table S4). In addition, risk factors strongly associated with risk in CONCORDANCE are also generally more common (medical management, CV event in the prior 6 months). Thus, in general, patients are more diverse in terms of the risk factors they present with, again facilitating easier discrimination.

| External validation
T A B L E 2 A refined predictive model for risk of the primary composite outcome and simplified scoring for the risk index The EQ-5D grades five dimensions (mobility, self-care, usual activities, pain/discomfort, anxiety/depression) as no, moderate, or severe problem. Scoring each as 0, 1, or 2 points, respectively, and adding these up yields an overall score ranging from 0-10. A score of 3 points means a patient had either: (a) three dimensions with moderate problem or (b) one dimension with moderate problem and one dimension with severe problem. A score of 4 or more points means a patient had at least either: (a) four dimensions with moderate problem; (b) two dimensions with moderate problem and one dimension with severe problem; (c) two dimensions with severe problem.    Independent contribution of increasing age is best captured by age over 65 years with increased risk represented by increased susceptibility to comorbidities and poorer quality of life included separately in the risk index.
Other concomitant conditions (diabetes, chronic kidney disease, heart failure, peripheral arterial disease and >1 MI) all contribute to increased risk. These likely reflect the greater burden of vascular disease in these patients. Having their MI medically managed only, and history of major bleeding also contributed to increased risk, possibly reflecting failure to tolerate or be offered prognostically important therapies. All of these factors have contributed to other risk scores in stable and unstable populations. 3,4,[16][17][18] A more innovative contributor to our risk index is patient selfreported health status, using a simple overall score derived from the EQ-5D-3L. Patients with poor self-reported health (3 points) had around a 50% increase in cardiovascular incidence rate, while those with very poor self-reported health (≥4 points) had around a doubling of incidence rate ( Table 2). This simple patient rating of health status was the strongest contributor to risk. The EQ-5D instrument is also reported as a strong predictor of mortality after discharge post-MI. 3,4 The reasons for this are speculative; in some patients the EQ-5D may be unmasking undetected depression, a known adverse marker of poor prognosis, 19 or it may be that patients with poorer self-reported quality of life are less likely to adhere to prescribed medications, or to attend cardiac rehabilitation, behaviors which have both been shown to adversely affect long-term outcomes 20,21 We also studied the UK-weighted index score for the EQ-5D-3L, revealing results comparable to our easier-to-use overall score.
For external validation, we used the CONCORDANCE registry 14 as it included all items in our risk index, and data on our primary outcome. This Australian population had follow-up starting 6 months after MI, so not a perfect match to our TIGRIS population. Nevertheless, the risk index achieves a broadly similar extent of risk discrimination for both the primary outcome and death. It is of considerable practical value to be able to identify which patients with stable coronary disease are at high risk of cardiovascular events and death. For instance, there are important risk reductions with new therapies for LDL-lowering 6 and relating to antithrombotic therapy. 7,8 Identifying the spectrum of risk in stable patients eligible for such treatments will identify those high-risk patients for whom the absolute reduction will be greatest.  (TRA 2 P-TIMI 50) trial, 23 a multivariable risk model was derived for cardiovascular death, MI, and ischemic stroke over a median 2.5 years in 8598 placebotreated patients recruited 2 weeks to 1-year post-MI. Their population is earlier post-MI, with 45% recruited within 3 months and 74% within 6 months. 24 Thus, early follow-up is in the post-acute phase when mortality is double that at 1-year post-MI. 25 Their risk model contains nine items, three of which (smoking, prior coronary artery bypass graft, and hypertension) were not independent predictors when applied to TIGRIS. Hence, applying the TIMI-50 risk model to our population showed weaker discrimination: c-statistic 0.63 and 0.70 for the primary outcome and all-cause death, respectively. However, we acknowledge that the c-statistic for our own risk score may be slightly optimistic, given that it is derived and assessed in the same population.
Battes et al 26  Regarding our study's limitations, based on 50 candidate predictors, there is risk of "false positives" entering our risk model. However, with P < .01 as entry criterion, risk is relatively low. The most novel highly significant predictor is EQ-5D overall score, while other risk index items are not surprising, having rational explanations based on prior studies. One problem potentially affecting the widespread use of our risk index may be that patients self-reported health status by EQ-5D is not routinely collected in most clinical settings. This is an emerging concern given the recognition of the importance of patient reported outcomes as indicators of the quality of care they receive. 29 In showing that patient reported outcomes makes a clear contribution to their prognosis, we provide additional justification for the importance of collection of data on patient symptom status and encourage a wider appreciation that an assessment of patient self-perception of their well-being is an important component of patient care.
While TIGRIS was designed to recruit representative patients in representative centers in each country, we cannot verify a truly generalizable population. Also, TIGRIS recruitment required each patient to have ≥1 of 5 risk criteria, four of which (all except multi-vessel coronary artery disease) are in the risk model. Thus, in applying our risk index to an unselected population of patients, a higher proportion may be identified as low risk.
We have successful external validation for our risk index using the CONCORDANCE registry and would encourage further validation studies in other relevant populations. While we studied patients rec-