Abstract
Background
Ruling out myocardial infarction (MI) in patients with an initial indeterminate (detectable to mildly elevated) troponin measure is challenging. Myocardial-Ischaemic-Injury Index (MI3) is a machine-learning algorithm designed to diagnose MI, but its utility in patients with indeterminate troponins is unclear. This study seeks to evaluate its diagnostic performance in patients with an initial indeterminate troponin.
Methods
We conducted a secondary analysis of a cohort (Cardiovascular Magnetic Resonance-Invasive-based Strategies in Patients with Chest Pain and Detectable to Mildly Elevated Serum Troponin) of adult patients with symptoms suggestive of acute coronary syndrome and an initial clinical contemporary troponin of 0.006–1.0 ng/mL across four US hospitals. Patients with initial and 3-hour high-sensitivity cardiac troponin I (Abbott Laboratories) measures were classified by MI3 into low-risk, intermediate-risk and high-risk groups. The primary outcome was adjudicated MI at 30 days. The sensitivity, specificity and negative likelihood ratio (−LR) of MI3 for MI at 30 days were calculated and reported with 95% CIs. A receiver operator characteristics curve for MI at 30 days was created and area under the curve (AUC) for MI3 was calculated.
Results
Among 207 patients, 34.3% (71/207) were female with a mean age of 61±11 years. MI at 30 days occurred in 43.5% (90/207). The AUC for MI3 for the detection of MI at 30 days was 0.882 (95% CI 0.833 to 0.932). MI3 classified 34.8% (72/207) of patients as low-risk, of which 8.3% (6/72) had MI at 30 days, yielding a sensitivity of 93.3% (95% CI 86.1 to 97.5%) and −LR of 0.12 (95% CI 0.05 to 0.26). Among the 47.3% (98/207) classified as intermediate-risk, MI at 30 days occurred in 48.0% (47/98). MI3 classified 17.9% (37/207) as high-risk, among which 100% (37/37) had MI at 30 days, yielding a specificity of 100% (95% CI 96.9% to 100%).
Conclusions
Among emergency department patients with an initial indeterminate troponin measure, the MI3 machine-learning algorithm had high AUC and specificity for 30-day MI.
Keywords: Acute Coronary Syndrome, RISK STRATIFICATION, Biomarkers
WHAT IS ALREADY KNOWN ON THIS TOPIC
The evaluation of patients with an indeterminate troponin measure is challenging and it is unclear whether machine-learning algorithms, like the Myocardial-Ischaemic-Injury Index (MI3) used alone or combined with other biomarkers, can aid the diagnosis of myocardial infarction in these patients.
WHAT THIS STUDY ADDS
In emergency department patients with acute chest pain and an initial indeterminate troponin measure, MI3 demonstrated excellent area under the curve and high specificity for 30-day myocardial infarction (MI), suggesting that it may be a useful diagnostic tool in these patients. However, MI3 was not sufficiently sensitive to exclude MI and the addition of B-type natriuretic peptide (BNP), Gal-3 and N-terminal pro BNP to MI3 did not substantively improve diagnostic performance.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
Our data suggest that machine-learning approaches to distinguishing MI from myocardial injury in patients with initial indeterminate high-sensitivity cardiac troponin measures hold promise and warrant further investigation.
Introduction
Each year, approximately 7–9 million patients with symptoms of possible acute coronary syndrome (ACS) present to US emergency departments (EDs).1 For the initial evaluation of these patients, guidelines recommend measurement of troponin to assess for myocardial injury and infarction.2 However, indeterminate (detectable or mildly elevated) results are common and are unable to rule in or rule out myocardial infarction (MI). Validated accelerated diagnostic protocols for patients with possible ACS risk stratify most of these patients into an ‘observation zone’ and recommend further investigation.3,5 Thus, the majority of these patients are admitted to a hospital ward or observation unit for further testing, such as stress testing or coronary angiography.5 However, the yield of these evaluations is relatively low, often with less than 20% ultimately diagnosed with MI.3,8
The Myocardial-Ischaemic-Injury Index (MI3) algorithm is a machine-learning algorithm which was developed and validated as an aid in diagnosis of type 1 MI.9 It uses patient’s age, sex and the value and timing of troponin measures to classify patients into low-risk, intermediate-risk and high-risk for MI and has previously demonstrated high diagnostic accuracy. Machine-learning algorithms, such as MI3, have the potential to improve risk stratification among patients with an indeterminate initial troponin, which could reduce low-yield hospitalisations and cardiac testing in this challenging patient population. However, the diagnostic performance of MI3 in ED patients with symptoms concerning for ACS and an initial indeterminate troponin is unknown. Furthermore, it is unclear whether MI3 performance can be improved when combined with other biomarkers, such as B-type natriuretic peptide (BNP), N-terminal pro BNP (NT-proBNP) or Galectin-3 (Gal-3).
This analysis seeks to address this significant evidence gap in distinguishing MI from other causes of myocardial injury within a multisite US cohort of ED patients with suspected ACS. The objective of this study was to evaluate whether the MI3 machine-learning algorithm used in patients with an initial indeterminate troponin can identify patients with an adjudicated diagnosis of MI. In addition, we sought to explore whether adding BNP, NT-proBNP or Gal-3 biomarker measures to MI3 could improve its diagnostic performance among these patients.
Methods
Study design and setting
A secondary analysis of the Cardiac MRI versuss Invasive-based Strategies in Patients with Chest Pain and Detectable to Mildly Elevated Serum Troponin Randomised Clinical Trial (CMR-IMPACT; ClinicalTrials.gov: NCT01931852) was conducted.10 CMR-IMPACT prospectively enrolled participants from four high-volume EDs at US tertiary care centres (Corewell Health William Beaumont University Hospital, formerly Beaumont Hospital, Royal Oak, MI; University of Mississippi, Jackson, MS; Ohio State University, Columbus, OH; Wake Forest University, Winston-Salem, NC) from September 2013 to July 2018. All participating institutions had cardiovascular magnetic resonance (CMR) stress testing and invasive angiography with revascularisation capabilities. Patients were randomised to either a CMR-based or invasive-based treatment strategy. Patients randomised to the invasive-based treatment arm were evaluated by the admitting or consulting cardiology team and underwent invasive angiography if the patient met American College of Cardiology/American Heart Association guideline recommendations. Participants randomised to the CMR-based strategy were similarly evaluated by the consulting cardiology or admitting team and underwent CMR stress imaging if deemed appropriate. The institutional review board at each study site approved this trial. Written informed consent was obtained from all participants in the study. The Standards for Reporting of Diagnostic Accuracy reporting guideline helped direct the research and manuscript development processes.
Population
Adults (≥21 years old) with symptoms suggestive of ACS and at least one contemporary troponin value above the limit of detection (LoD) with no troponin results >1.0 ng/mL were included. Troponin assays used by the individual sites included the Abbott Architect, Abbott iSTAT, Beckman Coulter Access, Roche Cobas, Siemens Advia Centaur and Siemens Vista.10 Study staff approached potentially eligible patients in the ED. Exclusion criteria included haemodynamic instability, new ST-segment elevation (≥1 mm) or depression (≥2 mm) at presentation, ongoing symptoms requiring emergent cardiac catheterisation, completion of index visit stress testing or coronary angiography prior to enrolment, life expectancy less than 6 months, prior severe multivessel coronary artery disease determined to be inappropriate for mechanical intervention, revascularisation in the past 6 months, prior solid organ transplant, creatinine clearance<30 mL/min or <60 mL/min with concurrent hepato-renal syndrome or chronic liver disease, and any other contraindications to CMR-stress imaging. In the primary analysis, patients with two high-sensitivity cardiac troponin I (hs-cTnI; Abbott Laboratories) measures were included (both the initial and 3-hour measures). Secondary analyses included patients with an initial hs-cTnI sample. The study flow diagram is shown in figure 1. A previous and separate secondary analysis of this cohort included only patients with a final adjudicated MI of type 1 or type 2.11 This differs from the current analysis which includes all patients with serial samples with or without MI of any type.
Figure 1. Study flow diagram. CMR-IMPACT, Cardiovascular Magnetic Resonance-Invasive-based Strategies in Patients with Chest Pain and Detectable to Mildly Elevated Serum Troponin; hs-cTnI, high-sensitivity cardiac troponin I; MI, myocardial infarction.
Biomarker analyses
Blood samples were collected at baseline (<1 hour from the initial clinical blood draw) and 3 hours later (±30 min) and EDTA plasma aliquots were stored at −70°C. hs-cTnI was measured using the Abbott Architect STAT hs-cTnI chemiluminescent microparticle immunoassay (Abbott Laboratories, Abbott Park, IL). This assay has a 99th percentile upper reference limit (URL) of 17 ng/L in women and 35 ng/L in men, LoD of 1.7 ng/L, and a coefficient of variation <10% at 4.7 ng/L. NT-proBNP was quantified using the NT-proBNP II electrochemiluminescence immunoassay (Roche Diagnostics, Indianapolis, IN). Galectin-3 and BNP were measured using Abbott Architect chemiluminescent microparticle immunoassays (Abbott Laboratories).
Myocardial-Ischaemic-Injury Index
The MI3 algorithm is a gradient boosting model which was previously developed and validated as an aid in the diagnosis and exclusion of MI.9 It incorporates age, sex, a baseline hs-cTnI concentration and the rate of change between baseline hs-cTnI and a subsequent (serial) hs-cTnI value to determine an index score corresponding to the probability of type 1 MI (ranging from 0 to 100). The MI3 algorithm was originally designed to evaluate specifically for the presence of type 1 MI in ED patients with symptoms concerning for ACS but was not specifically evaluated among patients with initial indeterminate troponin measures.
Outcomes
The primary outcome was MI at 30-days inclusive of the. Secondary outcomes included 30-day major adverse cardiac events (the composite of cardiac death, MI, or coronary revascularisation and 30-day coronary revascularisations). Study staff, who were blinded to MI3 hs-cTnI, BNP, NT-proBNP and galectin-3 results, reviewed patient medical records at 30 days using a structured data abstraction template to detect events occurring at that facility since the last follow-up period. Study staff then contacted participants by telephone at 30 to 90 days using a modified version of a previously described scripted follow-up dialogue to further clarify events since discharge, identify events occurring at other care facilities and determine healthcare utilisation since discharge.12
Adjudication occurred on all patients with a diagnosis of MI documented in the site’s electronic health record, an elevated local troponin measure, or death during the follow-up period and was conducted by at least two expert reviewers (in cardiology or emergency medicine) with a third reviewer serving as the tiebreaker in the event of disagreement. MI was identified using the fourth universal definition of MI: rise and fall of troponin (with at least one value ˃99th percentile URL) with symptoms of ischaemia, ECG evidence of ischaemia, a new regional wall motion abnormality on cardiac imaging, or identification of plaque rupture or erosion by coronary angiography.13 Cardiac death was determined based on the trial (Action to Control Cardiovascular Risk in Diabetes) definition.14 However, death due to stroke was not considered a cardiac death. Adjudicators had access to all relevant clinical information within the patient’s medical record to make these determinations, but were blinded to MI3 hs-cTnI, BNP, NT-proBNP and galectin-3 results.
Statistical analysis
Mean and SD were reported for continuous variables and number and percentage for categorical variables. The primary analysis population was restricted to patients with serial hs-cTnI samples available. In this population, receiver operating characteristic (ROC) curves were constructed for the continuous MI3 algorithm index score and individual biomarkers (BNP, Gal-3 and NTproBNP) for the 30-day outcomes of MI, major adverse cardiac events (MACE: the composite of cardiac death, MI or coronary revascularisation), and coronary revascularisation. Coronary revascularisation was defined by angioplasty with or without stenting or bypass grafting. Area under the curve (AUC) was then calculated for each ROC curve and reported along with a 95% CI. The MI3 algorithm was also evaluated at previously developed model cut points of <1.6 (low risk), 1.6 to <49.7 (intermediate risk) and ≥49.7 (high risk) to assess risk stratification for each outcome. These cut points were developed originally for rule-out and rule-in of index-visit type 1 MI specifically. Sensitivity, negative predictive value (NPV) and negative likelihood ratio (–LR) along with 95% CIs were presented for low-risk evaluations, and specificity, positive predictive value (PPV) and positive LR (+LR) with 95% CIs were presented for high-risk evaluations. Event rates with percentages were reported for intermediate-risk evaluations.
To assess if BNP, Gal-3 or NT-proBNP could add to the predictive power of MI3 for each outcome, logistic regression models were constructed adding BNP, Gal-3 and NT-proBNP concentrations at baseline as predictors with the MI3 index score. For each model, ROC curves were constructed and AUCs calculated with a 95% CI. Delong’s tests were used to compare the AUC between models incorporating additional biomarkers versus MI3 alone. As a supplemental evaluation, the same analyses for MI3 risk stratification and logistic regression with other biomarkers were performed on patients with baseline samples only using a baseline-only version of MI3 which has been previously described.15 A sensitivity analysis was conducted limiting the cohort to patients defined as having an indeterminate troponin based on the initial hs-cTn criteria recommended by the European Society of Cardiology.16 17 All significance testing was two-sided, and R vV.4.3.0 was used for all analyses (R Core Team, 2023).
Patient and public involvement
Patients or the public were not involved in the design, conduct, reporting or dissemination plans of this study or the parent trial.
Results
Among 207 patients with serial hs-cTnI samples, 34.3% (71/207) were female and 66.7% (138/207) were white with a mean age of 61±11 years. Within 30 days, 43.5% (90/270) experienced MI, 48.8% (101/207) had MACE and 27.1% (56/207) received coronary revascularisation. The cohort demographics are summarised in table 1. The study flow diagram is described in figure 1.
Table 1. Patient characteristics, overall and by 30-day MI status.
|
Patient characteristics |
Overall n (%) or mean (SD) (N=207) |
30-day MI n (%) or mean (SD) (N=90) |
No 30-day MI n (%) or mean (SD) (N=117) |
|---|---|---|---|
| Age, years | 60.5 (10.8) | 59.8 (10.5) | 61.1 (11.2) |
| Sex | |||
| Female | 71 (34.3) | 29 (32.2) | 42 (35.9) |
| Race | |||
| Asian | 2 (1.0) | 2 (2.2) | 0 (0) |
| Pacific Islander | 1 (0.5) | 1 (1.1) | 0 (0) |
| Black or African American | 64 (30.9) | 18 (20.0) | 46 (39.3) |
| White | 138 (66.7) | 67 (74.4) | 71 (60.7) |
| Other | 2 (1.0) | 2 (2.2) | 0 (0) |
| Ethnicity | |||
| Hispanic or Latino | 2 (1.0) | 2 (2.2) | 0 (0) |
| Risk factors | |||
| Weight, lbs | 200.7 (45.2) | 195.8 (44.7) | 204.5 (45.4) |
| Height, inches | 67.8 (4.2) | 67.9 (4.1) | 67.7 (4.3) |
| Current or history of smoking* | 133 (64.6) | 60 (66.7) | 73 (62.9) |
| Current or history of cocaine use* | 21 (10.2) | 8 (8.9) | 13 (11.3) |
| Hypertension | 160 (77.3) | 62 (68.9) | 98 (83.8) |
| Diabetes | 73 (35.3) | 25 (27.8) | 48 (41.0) |
| Hyperlipidaemia | 121 (58.5) | 49 (54.4) | 72 (61.5) |
| Prior congestive heart failure | 27 (13.0) | 11 (12.2) | 16 (13.7) |
| Prior coronary artery disease | 81 (39.1) | 32 (35.6) | 49 (41.9) |
| Prior MI | 52 (25.1) | 24 (26.7) | 28 (23.9) |
| Prior stent/PCI | 55 (26.6) | 24 (26.7) | 31 (26.5) |
| Prior CABG | 26 (12.6) | 10 (11.1) | 16 (13.7) |
| Prior coronary invasive angiography | 84 (40.6) | 35 (38.9) | 49 (41.9) |
| Prior cerebral vascular accident | 21 (10.1) | 9 (10.0) | 12 (10.3) |
| Prior peripheral vascular disease* | 24 (11.7) | 10 (11.2) | 14 (12.1) |
| Family history of ACS | 93 (44.9) | 40 (44.4) | 53 (45.3) |
Smoking is missing for one patient without MI, cocaine use is missing for 2 patients without MI, peripheral vascular disease is missing for one patient with MI and one patient without MI.
ACS, acute coronary syndrome; CABG, coronary artery bypass graft; MI, myocardial infarction; PCI, percutaneous coronary intervention.
The MI3 machine-learning algorithm classified 34.8% (72/270) of patients as low risk. Among these patients, 8.3% (6/72) had an MI within 30 days and 16.7% (12/72) had 30-day MACE, which included 6 patients with revascularisation procedures without adjudicated MI. This resulted in a sensitivity of 93.3% (95% CI 86.1% to 97.5%) for MI, 88.1% (95% CI 80.2% to 93.7%) for MACE and 89.3% (95% CI 78.1% to 96.0%) for revascularisation, respectively. MI3 yielded an NPV of 91.7% (95% CI 82.7% to 96.9%) and −LR of 0.12 (95% CI 0.05 to 0.26) for MI. NPV dropped to 83.3% (95% CI 72.7% to 91.1%) for MACE with an −LR of 0.21 (95% CI 0.12 to 0.37).
MI3 classified 17.9% (37/207) of the cohort as high risk. All high-risk patients (37/37) experienced a 30-day MI and MACE, resulting in a PPV of 100% (95% CI 90.5% to 100%) and specificity of 100% (95% CI 96.9% to 100% for MI, 95% CI 96.6% to 100% for MACE). Among these patients, 17.9% (37/207) also had revascularisation at 30 days, yielding a PPV of 73.0% (95% CI 55.9% to 86.2%), specificity of 93.4% (95% CI 88.2% to 96.8%) and +LR of 7.28 (95% CI 3.77 to 14.05) for revascularisation. Test characteristics for MI3 for MI, revascularisation and MACE at 30 days are summarised in table 2 and figure 2.
Table 2. Test characteristics of MI3 for 30-day MI, revascularisation and MACE.
| Test characteristics | 30-day MI | 30-day revascularisation | 30-day MACE |
|---|---|---|---|
| Low risk | |||
| Sensitivity (95% CI) | 93.3 (86.1 to 97.5) | 89.3 (78.1 to 96.0) | 88.1 (80.2 to 93.7) |
| NPV (95% CI) | 91.7 (82.7 to 96.9) | 91.7 (82.7 to 96.9) | 83.3 (72.7 to 91.1) |
| −LR (95% CI) | 0.12 (0.05 to 0.26) | 0.25 (0.11 to 0.53) | 0.21 (0.12 to 0.37) |
| High risk | |||
| Specificity (95% CI) | 100 (96.9 to 100) | 93.4 (88.2 to 96.8) | 100 (96.6 to 100) |
| PPV (95% CI) | 100 (90.5 to 100) | 73.0 (55.9 to 86.2) | 100 (90.5 to 100) |
| +LR (95% CI) | NA | 7.28 (3.77 to 14.05) | NA |
LR, likelihood ratio; +LR, positive LR; −LR, negative LR; MACE, major adverse cardiac events; MI, myocardial infarction; MI3, Myocardial-Ischaemic-Injury Index; NA, not available; NPV, negative predictive value; PPV, positive predictive value.
Figure 2. Risk stratification flow chart and test characteristics for MI3 for 30-day MI (A), coronary revascularisation (B), and MACE (C). MACE, major adverse cardiac event; MI, myocardial infarction; MI3, Myocardial-Ischaemic-Injury Index; NPV, negative predictive value; PPV, positive predictive value.
The MI3 algorithm had an AUC of 0.882 (95% CI 0.833 to 0.932) for the primary outcome of 30-day MI. This was higher than individual biomarkers BNP, Gal-3 and NT-proBNP with AUCs of 0.553 (95% CI 0.473 to 0.633), 0.555 (95% CI 0.476 to 0.634) and 0.597 (95% CI 0.476 to 0.634), respectively. The performance of the MI3 algorithm and individual biomarkers for the secondary outcomes of 30-day MACE and revascularisation had similarly discriminatory performance. The MI3 algorithm had an AUC of 0.854 (95% CI 0.801 to 0.907) for 30-day MACE and 0.818 (95% CI 0.751 to 0.885) for 30-day revascularisation events. The results of the AUC evaluation are summarised in table 3.
Table 3. AUC for MI3 and added biomarkers for 30-day MI, revascularisation and MACE.
| Marker | 30-day MI | 30-day revascularisation | 30-day MACE | |||
|---|---|---|---|---|---|---|
| AUC (95% CI) |
P value* | AUC (95% CI) |
P value* | AUC (95% CI) |
P value* | |
| MI3 | 0.882 (0.833 to 0.932) |
-- | 0.818 (0.751 to 0.885) |
-- | 0.854 (0.801 to 0.907) |
-- |
| BNP | 0.553 (0.473 to 0.633) |
-- | 0.493 (0.405 to 0.582) |
-- | 0.551 (0.472 to 0.630) |
-- |
| Gal-3 | 0.555 (0.476 to 0.634) |
-- | 0.634 (0.552 to 0.716) |
-- | 0.551 (0.472 to 0.630) |
-- |
| NT-proBNP | 0.597 (0.519 to 0.675) |
-- | 0.483 (0.395 to 0.570) |
-- | 0.591 (0.514 to 0.669) |
-- |
| MI3+BNP | 0.883 (0.836 to 0.931) |
0.708 | 0.815 (0.746 to 0.885) |
0.951 | 0.850 (0.795 to 0.904) |
0.669 |
| MI3+Gal-3 | 0.884 (0.835 to 0.932) |
0.874 | 0.823 (0.752 to 0.893) |
0.848 | 0.856 (0.803 to 0.908) |
0.866 |
| MI3+NT-proBNP | 0.883 (0.835 to 0.932) |
0.799 | 0.816 (0.746 to 0.887) |
0.904 | 0.852 (0.798 to 0.906) |
0.540 |
Compared with MI3 alone using Delong’s paired AUC test.
AUC, area under the curve; BNP, B-type natriuretic peptide; Gal-3, galectin-3; MACE, major adverse cardiac events; MI, myocardial infarction; MI3, Myocardial-Ischaemic-Injury Index; NT-proBNP, N-Terminal pro BNP.
Adding BNP, Gal-3 or NT-proBNP to MI3 demonstrated minimal benefit for 30-day MI, with no evidence of a statistically significant improvement in AUC (figure 3, table 3). These patterns were consistent with the evaluation of MI3 using only the baseline troponin result (online supplemental tables 1–3). Overall, MI3 had the strongest performance for 30-day MI, followed by MACE and revascularisation. Other biomarkers provided negligible improvement to the algorithm’s discriminative ability in this population. A sensitivity analysis limiting the cohort to those who meet the ESC definition of indeterminate initial hs-cTn did not substantively change the diagnostic performance of MI3 alone or in combination with BNP, Gal-3 or NT-proBNP (online supplemental materials).
Figure 3. ROC curves for 30-day MI using Gal-3 (A), BNP (B), NTproBNP (C), compared with MI3 alone and in combination with MI3. BNP, B-type natriuretic peptide; MI3, Myocardial-Ischaemic-Injury Index; NTproBNP, N-terminal pro BNP; ROC, receiver operating characteristic.
Discussion
Among ED patients with acute chest pain and an initial indeterminate troponin measure, the MI3 machine-learning algorithm had an excellent AUC and high specificity for 30-day MI, suggesting that it may be useful as an aid in diagnosis of MI in this challenging patient population. AUC and specificity were also high for 30-day MACE and revascularisation outcomes. However, MI3 was not sufficiently sensitive to exclude MI or MACE within this cohort. Finally, the addition of BNP, Gal-3 and NT-proBNP to MI3 did not substantively improve diagnostic performance for any 30-day outcome.
Currently, there are major gaps in the clinical differentiation of patients with MI versus other causes of acute myocardial injury among patients with mildly elevated hs-cTn measures.5 Validated accelerated diagnostic pathways, such as the European Society of Cardiology 0/1-hour algorithm, classify these patients to the ‘observation zone’ for further testing with additional serial hs-cTn and cardiac imaging. However, prior studies indicated that despite high rates of hospitalisations and cardiac testing in these patients, less than 20% are ultimately diagnosed with an MI.3,8 Therefore, development and validation of a tool to better risk stratify ED patients with indeterminate troponins has the potential to improve the quality and value of chest pain care by focusing downstream resources on patients that are more likely to have ACS.
Although the MI3 machine-learning algorithm was originally derived and validated to detect index visit type 1 MI among an all-comers population of ED patients with symptoms concerning for ACS,9 its application in patients with indeterminate hs-cTnI measures to detect a range of ACS (MI, MACE, coronary revascularisation) outcomes over 30 days was promising. MI3 had high AUC and specificity for 30-day MI, MACE and coronary revascularisation. These findings suggest that MI3 could be a useful tool to further risk stratify patients with indeterminate hs-cTn measures and that its utility for detecting MI extends beyond the index visit to 30 days. However, it appears to be better used as a rule-in tool, as its NPV, sensitivity and −LR for 30-day events are likely to be insufficient to allow for immediate ED discharge.
Prior studies suggest that biologically active BNP and its biologically inactive N-terminal fragment (NT-proBNP) may have added utility to hs-cTn in the diagnosis of MI. Furthermore, inflammatory biomarkers, such as galectin-3 (Gal-3), may be able to improve discrimination of downstream cardiac events, but have not been fully evaluated.18 19 However, in this analysis, the addition of BNP, NT-proBNP or Gal-3 to MI3 did not improve diagnostic accuracy for 30-day events. These findings are consistent with several studies that failed to demonstrate meaningful diagnostic improvement from adding other biomarkers to hs-cTn versus hs-cTn strategies used alone.20,24
As hospitals continue to transition to hs-cTn assays, the observed incidence of ED patients with initial indeterminate hs-cTn values is likely to increase, suggesting that the need for more accurate diagnostic approaches to differentiate MI from other causes of myocardial injury in this population. Thus, efforts to improve risk stratification in patients with indeterminate hs-cTn measures are critical. Our data suggest that a machine-learning approach holds promise in this challenging patient population. However, additional prospective multisite studies are needed to validate this strategy.
Limitations
Although the parent trial (CMR-IMPACT) was conducted across multiple US EDs, the sites were mostly urban academic medical centres and thus may not be generalisable to all ED settings. In addition, the parent trial screened many patients to identify the cohort, which could have introduced selection bias. The sample size was small, particularly for the subset of patients with serial hs-cTnI samples used in this analysis. Trial recruitment occurred prior to FDA approval of hs-cTn assays, thus hs-cTnI was measured from stored research samples and not assays used in clinical care at the time. Use of contemporary assays by the CMR IMPACT trial to define a population of patients with indeterminate troponin measures may have underestimated the number of MI events, as some patients classified as indeterminate risk based on CMR IMPACT trial criteria had elevated hs-cTnI measures. Furthermore, some patients classified in CMR IMPACT as having indeterminate troponin measures had non-elevated hs-cTnI measures from stored plasma. However, our sensitivity analysis limiting the cohort to patients meeting the ESC definition for an initial indeterminate hs-cTn did not substantively change our results. In addition, our results are from a secondary analysis of a trial that was not designed to assess the diagnostic performance of MI3 or cardiac biomarkers for cardiac outcomes. MI3 was originally derived to detect type 1 MI from an all-comers chest pain population not patients with indeterminate troponin or for additional cardiovascular outcomes. Also, in this analysis, MIs were not further characterised by MI type. This study focused on patients with initial indeterminate troponin measures and as such, this cohort is not representative of the broader ED population of patients with acute chest pain. Finally, while our data suggest that MI3 has high accuracy and specificity for 30-day MI in an ED population of patients with symptoms concerning for ACS and an initial indeterminate troponin, prospective multisite validation is needed.
Conclusions
In ED patients with symptoms concerning for ACS and an initial indeterminate troponin measure, the MI3 machine-learning algorithm had high AUC and specificity for 30-day MI, MACE and coronary revascularisation. These data suggest MI3 may be useful as an aid in diagnosis of MI versus other forms of myocardial injury in this challenging patient population. However, MI3 alone was not sensitive enough to exclude MI, MACE or need for revascularisation within this cohort. Further, the addition of natriuretic peptide measures (BNP and NT-proBNP) or Gal-3 to MI3 did not improve diagnostic performance for any outcome. While these data suggest that machine-learning approaches to distinguishing MI from myocardial injury in patients with initial indeterminate hs-cTn measures are promising, larger prospective studies are needed.
Supplementary material
Acknowledgements
The authors wish to thank Drs Subha V Raman, Jeffrey M Caterino, Carol L Clark, Alan E Jones, Michael E Hall, Lauren E Koehler, James F Lovato, Brian C Hiestand, Carolyn J Park, Sujethra Vasu, Michael A Kutcher and W Gregory Hundley for their work on the Cardiac Magnetic Resonance Imaging-IMPACT Trial.
Footnotes
Funding: This study was funded by Abbott Laboratories. Employees of the funder had a role in study design, data analysis and manuscript development. The parent study, the Cardiac Magnetic Resonance Imaging-IMPACT Trial, was funded by the National Heart Lung and Blood Institute (R01 HL118263-01).
Provenance and peer review: Not commissioned; externally peer reviewed.
Patient consent for publication: Not applicable.
Ethics approval: This study involves human participants and was approved by Wake Forest University School of Medicine Institutional Review Board (IRB00022322). Participants gave informed consent to participate in the study before taking part.
Data availability statement
Data are available on reasonable request.
References
- 1.Benjamin EJ, Blaha MJ, Chiuve SE, et al. Heart Disease and Stroke Statistics-2017 Update: A Report From the American Heart Association. Circulation. 2017;135:e146–603. doi: 10.1161/CIR.0000000000000485. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Gulati M, Levy PD, Mukherjee D, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation. 2021;144:e368–454. doi: 10.1161/CIR.0000000000001029. [DOI] [PubMed] [Google Scholar]
- 3.Reichlin T, Twerenbold R, Wildi K, et al. Prospective validation of a 1-hour algorithm to rule-out and rule-in acute myocardial infarction using a high-sensitivity cardiac troponin T assay. CMAJ. 2015;187:E243–52. doi: 10.1503/cmaj.141349. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Twerenbold R, Badertscher P, Boeddinghaus J, et al. 0/1-Hour Triage Algorithm for Myocardial Infarction in Patients With Renal Dysfunction. Circulation. 2018;137:436–51. doi: 10.1161/CIRCULATIONAHA.117.028901. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Nestelberger T, Wildi K, Boeddinghaus J, et al. Characterization of the observe zone of the ESC 2015 high-sensitivity cardiac troponin 0h/1h-algorithm for the early diagnosis of acute myocardial infarction. Int J Cardiol. 2016;207:238–45. doi: 10.1016/j.ijcard.2016.01.112. [DOI] [PubMed] [Google Scholar]
- 6.Nowak RM, Christenson RH, Jacobsen G, et al. Performance of Novel High-Sensitivity Cardiac Troponin I Assays for 0/1-Hour and 0/2- to 3-Hour Evaluations for Acute Myocardial Infarction: Results From the HIGH-US Study. Ann Emerg Med. 2020;76:1–13. doi: 10.1016/j.annemergmed.2019.12.008. [DOI] [PubMed] [Google Scholar]
- 7.Ashburn NP, Snavely AC, O’Neill JC, et al. Performance of the European Society of Cardiology 0/1-Hour Algorithm With High-Sensitivity Cardiac Troponin T Among Patients With Known Coronary Artery Disease. JAMA Cardiol. 2023;8:347–56. doi: 10.1001/jamacardio.2023.0031. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Allen BR, Christenson RH, Cohen SA, et al. Diagnostic Performance of High-Sensitivity Cardiac Troponin T Strategies and Clinical Variables in a Multisite US Cohort. Circulation. 2021;143:1659–72. doi: 10.1161/CIRCULATIONAHA.120.049298. [DOI] [PubMed] [Google Scholar]
- 9.Than MP, Pickering JW, Sandoval Y, et al. Machine Learning to Predict the Likelihood of Acute Myocardial Infarction. Circulation. 2019;140:899–909. doi: 10.1161/CIRCULATIONAHA.119.041980. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Miller CD, Mahler SA, Snavely AC, et al. Cardiac Magnetic Resonance Imaging Versus Invasive-Based Strategies in Patients With Chest Pain and Detectable to Mildly Elevated Serum Troponin: A Randomized Clinical Trial. Circ Cardiovasc Imaging. 2023;16:e015063. doi: 10.1161/CIRCIMAGING.122.015063. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Snavely A, Jackson L, Hunter CJ, et al. Differentiating Type 1 and Type 2 myocardial infarction using a machine learning algorithm and biomarkers. Am J Emerg Med. 2025;98:50–6. doi: 10.1016/j.ajem.2025.08.019. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Kline JA, Mitchell AM, Runyon MS, et al. Electronic medical record review as a surrogate to telephone follow-up to establish outcome for diagnostic research studies in the emergency department. Acad Emerg Med. 2005;12:1127–33. doi: 10.1197/j.aem.2005.04.012. [DOI] [PubMed] [Google Scholar]
- 13.Thygesen K, Alpert JS, Jaffe AS, et al. Fourth Universal Definition of Myocardial Infarction (2018) J Am Coll Cardiol. 2018;72:2231–64. doi: 10.1016/j.jacc.2018.08.1038. [DOI] [PubMed] [Google Scholar]
- 14.ACCORD Study Group. Cushman WC, Evans GW, et al. Effects of intensive blood-pressure control in type 2 diabetes mellitus. N Engl J Med. 2010;362:1575–85. doi: 10.1056/NEJMoa1001286. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Lin Z, Aw TC, Jackson L, et al. Machine learning to risk stratify chest pain patients with non-diagnostic electrocardiogram in an Asian emergency department. Ann Acad Med Singap. 2025;54:219–26. doi: 10.47102/annals-acadmedsg.2024232. [DOI] [PubMed] [Google Scholar]
- 16.Byrne RA, Rossello X, Coughlan JJ, et al. 2023 ESC Guidelines for the management of acute coronary syndromes. Eur Heart J. 2023;44:3720–826. doi: 10.1093/eurheartj/ehad191. [DOI] [PubMed] [Google Scholar]
- 17.Kontos MC, de Lemos JA, et al. 2022 ACC Expert Consensus Decision Pathway on the Evaluation and Disposition of Acute Chest Pain in the Emergency Department: A Report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol. 2022;80:1925–60. doi: 10.1016/j.jacc.2022.08.750. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Nowak RM, Jacobsen G, Christenson RH, et al. Differentiating type 1 and 2 acute myocardial infarctions using the N-terminal pro B-type natriuretic peptide/cardiac troponin T ratio. Am J Emerg Med. 2018;36:1849–54. doi: 10.1016/j.ajem.2018.06.073. [DOI] [PubMed] [Google Scholar]
- 19.Hara A, Niwa M, Kanayama T, et al. Galectin-3: A Potential Prognostic and Diagnostic Marker for Heart Disease and Detection of Early Stage Pathology. Biomolecules. 2020;10:1277. doi: 10.3390/biom10091277. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Yildirim M, Salbach C, Mueller-Hennessen M, et al. Diagnostic and Prognostic Evaluation of Novel Biomarkers Compared to ESC 0/1 h and 0/3 h Algorithms in Patients with Suspected Non-ST-Elevation Myocardial Infarction. J Clin Med. 2025;14:2957. doi: 10.3390/jcm14092957. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Karakas M, Januzzi JL, Jr, Meyer J, et al. Copeptin does not add diagnostic information to high-sensitivity troponin T in low- to intermediate-risk patients with acute chest pain: results from the rule out myocardial infarction by computed tomography (ROMICAT) study. Clin Chem. 2011;57:1137–45. doi: 10.1373/clinchem.2010.160192. [DOI] [PubMed] [Google Scholar]
- 22.Peacock WF, Nagurney J, Birkhahn R, et al. Myeloperoxidase in the diagnosis of acute coronary syndromes: the importance of spectrum. Am Heart J. 2011;162:893–9. doi: 10.1016/j.ahj.2011.08.017. [DOI] [PubMed] [Google Scholar]
- 23.Ashburn NP, Snavely AC, Allen BR, et al. Monocyte chemoattractant protein-1 is not predictive of cardiac events in patients with non-low-risk chest pain. Emerg Med J. 2022;39:853–8. doi: 10.1136/emermed-2021-211266. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Poldervaart JM, Röttger E, Dekker MS, et al. No Added Value of Novel Biomarkers in the Diagnostic Assessment of Patients Suspected of Acute Coronary Syndrome. PLoS One. 2015;10:e0132000. doi: 10.1371/journal.pone.0132000. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data are available on reasonable request.