https://orcid.org/0000-0002-9706-7900
Abstract
Our goal was to evaluate a previously validated artificial intelligence–augmented electrocardiography (AI-ECG) screening tool for left ventricular systolic dysfunction (LVSD) in patients undergoing high-sensitivity-cardiac troponin T (hs-cTnT).
Retrospective application of AI-ECG for LVSD in emergency department (ED) patients undergoing hs-cTnT. AI-ECG scores (0–1) for probability of LVSD (left ventricular ejection fraction ≤ 35%) were obtained. An AI-ECG score ≥0.256 indicates a positive screen. The primary endpoint was a composite of post-discharge major adverse cardiovascular events (MACEs) at two years follow-up. Among 1977 patients, 248 (13%) had a positive AI-ECG. When compared with patients with a negative AI-ECG, those with a positive AI-ECG had a higher risk for MACE [48 vs. 21%, P < 0.0001, adjusted hazard ratio (HR) 1.39, 95% confidence interval (CI) 1.11–1.75]. This was largely because of a higher rate of deaths (32 vs. 14%, P < 0.0001; adjusted HR 1.26, 95% 0.95–1.66) and heart failure hospitalizations (26 vs. 6.1%, P < 0.001; adjusted HR 1.75, 95% CI 1.25–2.45). Together, hs-cTnT and AI-ECG resulted in the following MACE rates and adjusted HRs: hs-cTnT < 99th percentile and negative AI-ECG: 116/1176 (11%; reference), hs-cTnT < 99th percentile and positive AI-ECG: 28/107 (26%; adjusted HR 1.54, 95% CI 1.01–2.36), hs-cTnT > 99th percentile and negative AI-ECG: 233/553 (42%; adjusted HR 2.12, 95% CI 1.66, 2.70), and hs-cTnT > 99th percentile and positive AI-ECG: 91/141 (65%; adjusted HR 2.83, 95% CI 2.06, 3.87).
Among ED patients evaluated with hs-cTnT, a positive AI-ECG for LVSD identifies patients at high risk for MACE. The conjoint use of hs-cTnT and AI-ECG facilitates risk stratification.
AI-ECG screen for left ventricular dysfunction among ED patients undergoing hs-cTnT measurements. Frequency of a positive AI-ECG screen for LVSD in ED patients undergoing hs-cTnT (upper left panel) and its prognostic implications in terms of MACE at 2 years after the index presentation (upper right panel). Lower panel: conjoint use of hs-cTnT and AI-ECG screen for risk stratification in the ED for 2-year MACE (on the left) and mortality (on the right) with adjusted HR for each combination.
Introduction
High-sensitivity cardiac troponin (hs-cTn) I and T assays are the preferred biomarkers for the detection of myocardial injury and to support the diagnosis of acute myocardial infarction (MI).1 These assays allow the rapid risk stratification of patients with suspected acute coronary syndrome (ACS).2,3 With hs-cTn assays, however, there can be substantial increases in the proportions of patients with hs-cTn concentrations above the 99th percentile upper-reference limit (URL).4 Although some of these patients will have Type 1 MI and be managed following clinical practice guideline recommendations, most patients with hs-cTn increases will have some form of acute or chronic non-ischaemic myocardial injury or Type 2 MI that is often associated with established or previously unrecognized underlying heart disease.1 These patients are at high risk for major adverse cardiovascular events, among which heart failure is increasingly recognized as major contributor.5–8 However, when compared with Type 1 MI, these groups are less likely to undergo cardiac evaluations and perhaps for this reason, they are often under-treated,4,5,9–11 and how to optimally identify those at highest risk for adverse events remains unclear.
It is important to identify patients with left ventricular systolic dysfunction (LVSD) as this is a serious condition associated with an adverse prognosis. Importantly, recognition of LVSD should lead to additional evaluations and initiation of evidence-based therapies to improve outcomes.12,13 The 2021 AHA/ACC clinical practice guidelines for the evaluation and diagnosis of chest pain3 recommend echocardiography for intermediate-risk patients to establish baseline ventricular function and indicate that patients with suspected ACS identified to have new-onset LVSD should be designated as high risk. It is challenging, however, and often not feasible nor cost-effective in many clinical contexts, to perform echocardiographic evaluations in all such patients, particularly in those with myocardial injury that are increasingly encountered using hs-cTn. In addition, this strategy has the potential to delay evaluations and exacerbate emergency department (ED) and hospital overcrowding with its attendant consequences. Several studies demonstrate that often less than half of patients with Type 2 MI or myocardial injury undergo echocardiography.4,5,14,15 For these reasons, novel, pragmatic, and cost-effective approaches to identifying LVSD in every patient would be valuable.
A novel artificial intelligence–augmented electrocardiogram (AI-ECG) algorithm was previously developed and extensively validated to recognize electrocardiographic patterns indicative of LVSD from a standard 12-lead ECG.16,17 It has been validated in ED patients with dyspnoea18 and in critically ill patients from the cardiac intensive care unit.19 AI-ECG may be a useful screening tool for the rapid and inexpensive risk stratification of LVSD. It may facilitate the identification of high-risk patients in whom additional evaluations are warranted, and low-risk patients in whom additional testing can be deferred or avoided. The goal of our study was to investigate the clinical application of this previously developed and validated AI-ECG for LVSD in ED patients undergoing hs-cTnT measurement.
Methods
Study design and patient population
The MAyo Southwest WisConsin 5th Gen Troponin T ImplementatiON (ACTION) study is a retrospective, multicenter (n = 2), observational cohort study of consecutive encounters of adult ED patients undergoing at least 1 cTnT measurement on clinical indication from 12 March 2018 to 11 March 2019 during the transition from 4th Gen cTnT to 5th Gen cTnT across the Southwest Wisconsin Mayo Clinic Health System (MCHS) hospitals at La Crosse and Sparta in Wisconsin. It excluded patients who did not present through the ED, were <18 years old, or had both 4th and 5th Gen cTnT measurements at transition. The study design and primary results have been published.4 This study was approved by the Institutional Review Board (ID 19-002668). Data were abstracted and reviewed from the electronic health records by trained staff following a standardized data collection process in Research Electronic Data Capture (REDCap).20 For the present analyses, we addressed unique patients based on their first presentation during the hs-cTnT study period. Patients aged <18 years old, who did not present through the ED, who had fourth generation cTnT measured, those without a 12-lead ECG, those in whom hs-cTn was not measured within 12 h of ED presentation, without an available AI-ECG probability of LVSD based on the index ECG, and those adjudicated as having Type 3–5 MIs were excluded (see Supplementary material online, Figure S1).
High-sensitivity-cardiac troponin T measurements
Hs-cTnT was measured using the Elecsys Troponin T Gen 5 STAT assay (Roche Diagnostics) on the Cobas e 601 (MCHS La Crosse) and Cobas e 411 (MCHS Sparta). Concentrations are reported as whole units (no decimals) in ng/L and results reported down to the limit of quantitation of <6 ng/L. Sex-specific 99th percentile URLs of 10 ng/L for women and 15 ng/L for men were used.21 Concentrations >10 ng/L for women and >15 ng/L for men are considered indicative of myocardial injury.1 The rationale for the Mayo Clinic 0 and 2 h hs-cTnT protocol for the rule-in and rule-out of MI has been described21 as have the results of its implementation.4
Myocardial injury and infarction adjudication
All available data from the clinical presentation during index hospitalization, including 12-lead ECG, transthoracic echocardiography (TTE), stress test, and angiograms, were reviewed. All encounters with at least 1 hs-cTnT >99th percentile URLs were adjudicated using the Fourth Universal Definition of Myocardial Infarction (UDMI) criteria1 by trained physicians. Cases with challenging adjudication were reviewed by the principal investigator (Y.S.), and if needed, by a member of the Task Force for the Fourth UDMI1 (A.S.J.). Patients with at least 1 hs-cTnT concentration >99th percentile URLs were classified as having either myocardial injury or acute MI if there were clinical features acute myocardial ischaemia, such as ischaemic symptoms, new ischaemic electrocardiogram changes, development of pathological Q waves, imaging evidence of new loss of viable myocardium or new regional wall motion abnormality in a pattern consistent with an ischaemic aetiology, and/or identification of a culprit lesion on coronary angiography.1 Those with overt evidence of myocardial ischaemia were classified as having MI and further subclassified into Type 1 or 2 MI subtypes1 (Supplemental material online, Methods). The term ‘myocardial injury’ applies to patients with hs-cTnT increases above the 99th percentile without clinical evidence of acute myocardial ischaemia.
Artificial intelligence–augmented electrocardiography algorithm for left ventricular systolic dysfunction
The derivation and validation of the AI-ECG algorithm for detection of LVSD have been described.16,17 The present study aims to evaluate its retrospective application in ED patients undergoing hs-cTnT measurement. In brief, a convolutional neural network was trained to detect LVSD based on a left ventricular ejection fraction (LVEF) ≤35% by TTE using data from ECGs and TTE from nearly 100 000 patients at the Mayo Clinic.16 AI-ECG probabilities for LVSD are reported for clinical use within the electronic health record in the Mayo Clinic AI-ECG Dashboard and are available for all digitalized 12-lead ECGs. Probabilities are reported as a continuous result between 0 and 1 that LVSD (LVEF ≤ 35%) is present based on the surface ECG. To identify patients at high probability of LVSD, a previously determined screening threshold of ≥0.256 to indicate a positive AI-ECG screen has been validated.17,18 Conversely, if the probability of LVSD is <0.256, then the AI-ECG screen is considered negative. We collected the AI-ECG probabilities estimated from the first 12-lead ECG obtained during the index presentation and entered them into REDCap. As a secondary analysis, among patients with an available TTE at index presentation or within 30 or 180 days post discharge, we also examined the diagnostic performance of the AI-ECG algorithm for the identification of LVSD. We also evaluated the prognostic significance of an adjudicated ischaemic 12-lead ECG (ischaemic ST elevation of ST-depression or T-wave inversion).
Study endpoints
The primary endpoint was post-discharge major adverse cardiovascular events (MACEs), which was a composite of all-cause mortality, acute MI, heart failure hospitalization, stroke or transient ischaemic attack, and new-onset atrial fibrillation or flutter up to 2 years follow-up.
Statistical analysis
Baseline characteristics are reported using median (interquartile range) and compared using the Kruskal–Wallis test for continuous variables. Categorical variables are presented as numbers (percentages) and compared using the χ2 test. Hazard ratios (HRs) and corresponding 95% confidence intervals (CIs) are presented. Multivariable models were developed including age, sex, heart failure, chronic kidney disease, and ischaemic heart disease. An expanded multivariable model incorporated adjudicated ECG abnormalities (ischaemic/non-ischaemic and sinus/non-sinus rhythm). Outcome analyses evaluated AI-ECG alone and AI-ECG together with hs-cTnT results. A P-value <0.05 was considered statistically significant. All analyses were performed using SAS version 9.4 (Cary, NC, USA) and R version 4.0.2.
Results
The study cohort included 1977 patients. The mean age of the population was 62 years, 52% were women, and 50% reported chest discomfort at presentation. Among the 1977 patients, 847 (43%) had at least one hs-cTnT concentration above the sex-specific 99th percentile. In these patients, 694 (82%) were adjudicated as having myocardial injury, 63 (7%) were classified as Type 1 MI, and 90 (11%) as Type 2 MI.
Baseline characteristics according to artificial intelligence–augmented electrocardiography screen for left ventricular systolic dysfunction
Baseline characteristics are shown in Table 1. A total of 1729 (87%) patients had a negative AI-ECG screen, while 248 (13%) had a positive AI-ECG screen for LVSD (Graphical abstract). The distribution of AI-ECG probabilities for LVSD is shown in Supplementary material online, Figure S2. Among patients with a negative AI-ECG, the median AI-ECG probabilities were 1.3 (0.5–3.0), whereas in those with a positive AI-ECG, the median AI-ECG probabilities were 68.1 (44.9–90.1; P < 0.0001; see Supplementary material online, Table S1). When compared with patients with a negative AI-ECG screen, patients with a positive AI-ECG screen were older, more frequently men, more likely to present with dyspnoea, and with more comorbidities (Table 1). They also had higher baseline and maximum median hs-cTnT concentrations and were more likely to have ST-T-wave changes on the presenting 12-lead ECG.
Baseline characteristics in the overall population and by negative or positive artificial intelligence–augmented electrocardiography screen
| Entire cohort | Negative AI-ECG screen (<0.256) | Positive AI-ECG screen (≥0.256) | P-value | |
|---|---|---|---|---|
| n = 1977 | n = 1729 | n = 248 | ||
| Demographics | ||||
| Age, mean (SD) | 62 (18) | 61 (18) | 70 (16) | <0.0001 |
| Women, n (%) | 1022 (52) | 942 (55) | 80 (32) | <0.0001 |
| Presenting symptoms | ||||
| Chest discomfort, n (%) | 990 (50) | 902 (52) | 88 (36) | <0.0001 |
| Dyspnoea, n (%) | 820 (42) | 682 (39) | 138 (56) | <0.0001 |
| Comorbidities | ||||
| Ischaemic heart disease, n (%) | 437 (22) | 315 (18) | 122 (49) | <0.0001 |
| Hypertension, n (%) | 1144 (58) | 964 (56) | 180 (73) | <0.0001 |
| Obesity, n (%) | 848 (43) | 745 (43) | 103 (42) | 0.64 |
| Current or prior tobacco use, n (%) | 1106 (56) | 956 (55) | 150 (61) | 0.12 |
| Cerebrovascular disease, n (%) | 198(10) | 166 (10) | 32 (13) | 0.11 |
| History of atrial fibrillation/flutter/tachycardia, n (%) | 371 (19) | 267 (15) | 104 (42) | <0.0001 |
| Heart failure, n (%) | 373 (19) | 236 (14) | 137 (55) | <0.0001 |
| Diabetes mellitus, n (%) | 477 (24) | 380 (22) | 97 (39) | <0.0001 |
| Chronic kidney disease, n (%) | 398 (20) | 296 (17) | 102 (41) | <0.0001 |
| Family history of CAD, n (%) | 614 (31) | 539 (31) | 75 (30) | 0.77 |
| Peripheral artery disease, n (%) | 204 (10) | 158 (9) | 46 (19) | <0.0001 |
| Dyslipidaemia, n (%) | 1022 (52) | 855 (50) | 167 (67) | <0.0001 |
| Laboratory results | ||||
| Median baseline hs-cTnT (ng/L) | 10 (0–21) | 9 (0–19) | 24 (13–49) | <0.0001 |
| Median maximum hs-cTnT (ng/L) | 10 (0–23) | 9 (0–20) | 28 (14–72) | <0.0001 |
| Creatinine (mg/dL), n (%) (mg/dL) (n = 1960) | 0.9 (0.8–1.1) | 0.9 (0.7–1.1) | 1.1 (0.9–1.5) | <0.0001 |
| eGFR, n (%) (mL/min/1.73 m2) (n = 1959) | 80 (59–91) | 82 (63–91) | 61 (39–86) | <0.0001 |
| NT-proBNP, n (%) (pg/mL) (n = 709) | 412 (94–2118) | 244 (77–1172) | 2484 (639–6231) | <0.0001 |
| Initial 12-lead electrocardiogram | ||||
| Sinus rhythm, n (%) | 1657 (84) | 1517 (88) | 140 (57) | <0.0001 |
| Non-ischaemic, n (%) | 1825 (92) | 1625 (94) | 200 (81) | <0.0001 |
| ST elevation, n (%) | 23 (1.2) | 11 (0.6) | 12 (4.8) | <0.0001 |
| ST depression or T-wave inversion, n (%) | 129 (6.5) | 93 (5.4) | 36 (15) | <0.0001 |
| Entire cohort | Negative AI-ECG screen (<0.256) | Positive AI-ECG screen (≥0.256) | P-value | |
|---|---|---|---|---|
| n = 1977 | n = 1729 | n = 248 | ||
| Demographics | ||||
| Age, mean (SD) | 62 (18) | 61 (18) | 70 (16) | <0.0001 |
| Women, n (%) | 1022 (52) | 942 (55) | 80 (32) | <0.0001 |
| Presenting symptoms | ||||
| Chest discomfort, n (%) | 990 (50) | 902 (52) | 88 (36) | <0.0001 |
| Dyspnoea, n (%) | 820 (42) | 682 (39) | 138 (56) | <0.0001 |
| Comorbidities | ||||
| Ischaemic heart disease, n (%) | 437 (22) | 315 (18) | 122 (49) | <0.0001 |
| Hypertension, n (%) | 1144 (58) | 964 (56) | 180 (73) | <0.0001 |
| Obesity, n (%) | 848 (43) | 745 (43) | 103 (42) | 0.64 |
| Current or prior tobacco use, n (%) | 1106 (56) | 956 (55) | 150 (61) | 0.12 |
| Cerebrovascular disease, n (%) | 198(10) | 166 (10) | 32 (13) | 0.11 |
| History of atrial fibrillation/flutter/tachycardia, n (%) | 371 (19) | 267 (15) | 104 (42) | <0.0001 |
| Heart failure, n (%) | 373 (19) | 236 (14) | 137 (55) | <0.0001 |
| Diabetes mellitus, n (%) | 477 (24) | 380 (22) | 97 (39) | <0.0001 |
| Chronic kidney disease, n (%) | 398 (20) | 296 (17) | 102 (41) | <0.0001 |
| Family history of CAD, n (%) | 614 (31) | 539 (31) | 75 (30) | 0.77 |
| Peripheral artery disease, n (%) | 204 (10) | 158 (9) | 46 (19) | <0.0001 |
| Dyslipidaemia, n (%) | 1022 (52) | 855 (50) | 167 (67) | <0.0001 |
| Laboratory results | ||||
| Median baseline hs-cTnT (ng/L) | 10 (0–21) | 9 (0–19) | 24 (13–49) | <0.0001 |
| Median maximum hs-cTnT (ng/L) | 10 (0–23) | 9 (0–20) | 28 (14–72) | <0.0001 |
| Creatinine (mg/dL), n (%) (mg/dL) (n = 1960) | 0.9 (0.8–1.1) | 0.9 (0.7–1.1) | 1.1 (0.9–1.5) | <0.0001 |
| eGFR, n (%) (mL/min/1.73 m2) (n = 1959) | 80 (59–91) | 82 (63–91) | 61 (39–86) | <0.0001 |
| NT-proBNP, n (%) (pg/mL) (n = 709) | 412 (94–2118) | 244 (77–1172) | 2484 (639–6231) | <0.0001 |
| Initial 12-lead electrocardiogram | ||||
| Sinus rhythm, n (%) | 1657 (84) | 1517 (88) | 140 (57) | <0.0001 |
| Non-ischaemic, n (%) | 1825 (92) | 1625 (94) | 200 (81) | <0.0001 |
| ST elevation, n (%) | 23 (1.2) | 11 (0.6) | 12 (4.8) | <0.0001 |
| ST depression or T-wave inversion, n (%) | 129 (6.5) | 93 (5.4) | 36 (15) | <0.0001 |
Values are mean (SD) or n (%). Laboratory results are median (interquartile ranges).
CAD, coronary artery disease; eGFR, estimated glomerular filtration rate; hs-cTnT, high-sensitivity-cardiac troponin T; MI, myocardial infarction.
Baseline characteristics in the overall population and by negative or positive artificial intelligence–augmented electrocardiography screen
| Entire cohort | Negative AI-ECG screen (<0.256) | Positive AI-ECG screen (≥0.256) | P-value | |
|---|---|---|---|---|
| n = 1977 | n = 1729 | n = 248 | ||
| Demographics | ||||
| Age, mean (SD) | 62 (18) | 61 (18) | 70 (16) | <0.0001 |
| Women, n (%) | 1022 (52) | 942 (55) | 80 (32) | <0.0001 |
| Presenting symptoms | ||||
| Chest discomfort, n (%) | 990 (50) | 902 (52) | 88 (36) | <0.0001 |
| Dyspnoea, n (%) | 820 (42) | 682 (39) | 138 (56) | <0.0001 |
| Comorbidities | ||||
| Ischaemic heart disease, n (%) | 437 (22) | 315 (18) | 122 (49) | <0.0001 |
| Hypertension, n (%) | 1144 (58) | 964 (56) | 180 (73) | <0.0001 |
| Obesity, n (%) | 848 (43) | 745 (43) | 103 (42) | 0.64 |
| Current or prior tobacco use, n (%) | 1106 (56) | 956 (55) | 150 (61) | 0.12 |
| Cerebrovascular disease, n (%) | 198(10) | 166 (10) | 32 (13) | 0.11 |
| History of atrial fibrillation/flutter/tachycardia, n (%) | 371 (19) | 267 (15) | 104 (42) | <0.0001 |
| Heart failure, n (%) | 373 (19) | 236 (14) | 137 (55) | <0.0001 |
| Diabetes mellitus, n (%) | 477 (24) | 380 (22) | 97 (39) | <0.0001 |
| Chronic kidney disease, n (%) | 398 (20) | 296 (17) | 102 (41) | <0.0001 |
| Family history of CAD, n (%) | 614 (31) | 539 (31) | 75 (30) | 0.77 |
| Peripheral artery disease, n (%) | 204 (10) | 158 (9) | 46 (19) | <0.0001 |
| Dyslipidaemia, n (%) | 1022 (52) | 855 (50) | 167 (67) | <0.0001 |
| Laboratory results | ||||
| Median baseline hs-cTnT (ng/L) | 10 (0–21) | 9 (0–19) | 24 (13–49) | <0.0001 |
| Median maximum hs-cTnT (ng/L) | 10 (0–23) | 9 (0–20) | 28 (14–72) | <0.0001 |
| Creatinine (mg/dL), n (%) (mg/dL) (n = 1960) | 0.9 (0.8–1.1) | 0.9 (0.7–1.1) | 1.1 (0.9–1.5) | <0.0001 |
| eGFR, n (%) (mL/min/1.73 m2) (n = 1959) | 80 (59–91) | 82 (63–91) | 61 (39–86) | <0.0001 |
| NT-proBNP, n (%) (pg/mL) (n = 709) | 412 (94–2118) | 244 (77–1172) | 2484 (639–6231) | <0.0001 |
| Initial 12-lead electrocardiogram | ||||
| Sinus rhythm, n (%) | 1657 (84) | 1517 (88) | 140 (57) | <0.0001 |
| Non-ischaemic, n (%) | 1825 (92) | 1625 (94) | 200 (81) | <0.0001 |
| ST elevation, n (%) | 23 (1.2) | 11 (0.6) | 12 (4.8) | <0.0001 |
| ST depression or T-wave inversion, n (%) | 129 (6.5) | 93 (5.4) | 36 (15) | <0.0001 |
| Entire cohort | Negative AI-ECG screen (<0.256) | Positive AI-ECG screen (≥0.256) | P-value | |
|---|---|---|---|---|
| n = 1977 | n = 1729 | n = 248 | ||
| Demographics | ||||
| Age, mean (SD) | 62 (18) | 61 (18) | 70 (16) | <0.0001 |
| Women, n (%) | 1022 (52) | 942 (55) | 80 (32) | <0.0001 |
| Presenting symptoms | ||||
| Chest discomfort, n (%) | 990 (50) | 902 (52) | 88 (36) | <0.0001 |
| Dyspnoea, n (%) | 820 (42) | 682 (39) | 138 (56) | <0.0001 |
| Comorbidities | ||||
| Ischaemic heart disease, n (%) | 437 (22) | 315 (18) | 122 (49) | <0.0001 |
| Hypertension, n (%) | 1144 (58) | 964 (56) | 180 (73) | <0.0001 |
| Obesity, n (%) | 848 (43) | 745 (43) | 103 (42) | 0.64 |
| Current or prior tobacco use, n (%) | 1106 (56) | 956 (55) | 150 (61) | 0.12 |
| Cerebrovascular disease, n (%) | 198(10) | 166 (10) | 32 (13) | 0.11 |
| History of atrial fibrillation/flutter/tachycardia, n (%) | 371 (19) | 267 (15) | 104 (42) | <0.0001 |
| Heart failure, n (%) | 373 (19) | 236 (14) | 137 (55) | <0.0001 |
| Diabetes mellitus, n (%) | 477 (24) | 380 (22) | 97 (39) | <0.0001 |
| Chronic kidney disease, n (%) | 398 (20) | 296 (17) | 102 (41) | <0.0001 |
| Family history of CAD, n (%) | 614 (31) | 539 (31) | 75 (30) | 0.77 |
| Peripheral artery disease, n (%) | 204 (10) | 158 (9) | 46 (19) | <0.0001 |
| Dyslipidaemia, n (%) | 1022 (52) | 855 (50) | 167 (67) | <0.0001 |
| Laboratory results | ||||
| Median baseline hs-cTnT (ng/L) | 10 (0–21) | 9 (0–19) | 24 (13–49) | <0.0001 |
| Median maximum hs-cTnT (ng/L) | 10 (0–23) | 9 (0–20) | 28 (14–72) | <0.0001 |
| Creatinine (mg/dL), n (%) (mg/dL) (n = 1960) | 0.9 (0.8–1.1) | 0.9 (0.7–1.1) | 1.1 (0.9–1.5) | <0.0001 |
| eGFR, n (%) (mL/min/1.73 m2) (n = 1959) | 80 (59–91) | 82 (63–91) | 61 (39–86) | <0.0001 |
| NT-proBNP, n (%) (pg/mL) (n = 709) | 412 (94–2118) | 244 (77–1172) | 2484 (639–6231) | <0.0001 |
| Initial 12-lead electrocardiogram | ||||
| Sinus rhythm, n (%) | 1657 (84) | 1517 (88) | 140 (57) | <0.0001 |
| Non-ischaemic, n (%) | 1825 (92) | 1625 (94) | 200 (81) | <0.0001 |
| ST elevation, n (%) | 23 (1.2) | 11 (0.6) | 12 (4.8) | <0.0001 |
| ST depression or T-wave inversion, n (%) | 129 (6.5) | 93 (5.4) | 36 (15) | <0.0001 |
Values are mean (SD) or n (%). Laboratory results are median (interquartile ranges).
CAD, coronary artery disease; eGFR, estimated glomerular filtration rate; hs-cTnT, high-sensitivity-cardiac troponin T; MI, myocardial infarction.
Artificial intelligence–augmented electrocardiography for left ventricular systolic dysfunction based on high-sensitivity-cardiac troponin T results stratified according to sex-specific 99th percentiles
Patients with increased hs-cTnT >99th percentile were more likely to have a positive AI-ECG than those without hs-cTnT increases (22 vs. 5.8%, P < 0.0001). Likewise, among patients with a positive AI-ECG, 183 (74%) had at least one hs-cTnT above the 99th percentile, whereas among patients with a negative AI-ECG, 664 (38%) had hs-cTnT increases (P < 0.0001). Median AI-ECG probabilities according to hs-cTnT results are shown in Supplementary material online, Table S1. Patients with increased hs-cTnT had higher AI-ECG probabilities for LVSD than those without hs-cTnT increases. Among patients with a positive AI-ECG screen, those with at least one hs-cTnT increase above the 99th percentile had higher median probabilities for LVSD than those without increases (74.1 vs. 52.5, P = 0.0002).
Artificial intelligence–augmented electrocardiography for left ventricular systolic dysfunction based on adjudicated diagnoses according to the Fourth Universal Definition of Myocardial Infarction
Patients with Type 1 MI had the highest proportion of a positive AI-ECG screen (38%), when compared with patients with Type 2 MI and myocardial injury in whom the frequency of a positive AI-ECG was 20% in both groups. Median AI-ECG probabilities according to adjudicated diagnoses are shown in Supplementary material online, Table S1. Patients with Type 1 MI had the highest AI-ECG probabilities for LVSD, followed by Type 2 MI and myocardial injury. Among patients with a positive AI-ECG screen, patients with Type 1 and 2 MIs have similar median AI probabilities of 85.3 and 85.8 respectively, which were higher than those of 71.9 observed in patients with myocardial injury (see Supplementary material online, Table S1).
Diagnostic performance of artificial intelligence–augmented electrocardiography for left ventricular systolic dysfunction among patients with available transthoracic echocardiography
A significant correlation was observed between the probability of LVEF ≤ 35% by AI-ECG algorithm values and LVEF among patients with an available echocardiogram (r = −0.474, P < 0.0001; see Supplementary material online, Figure S3). For patients with an available TTE during index presentation or within 30 days post discharge (n = 452), the diagnostic accuracy of the AI-ECG for LVSD was 81.4% (95% CI 77.5–84.9) with a specificity of 82.1% (95% CI 78.0–85.8) and sensitivity of 75.5% (95% CI 61.1–86.7). Similar results were obtained among 562 patients with a TTE during index presentation or within 180 days post discharge, with a diagnostic accuracy of 82.7% (95% CI 79.4–85.8; see Supplementary material online, Table S2).
The proportion of patients with LVEF ≤35, 35–50, and ≥50% based on TTE at index presentation or within 30 days post discharge in those with a positive and negative AI-ECG screen is shown in Figure 1. Most patients (297/343, 87%) with a negative AI-ECG screen had LVEF > 50%. The AI-ECG negative predictive values for LVEF ≤ 35% were 96.5% (95% CI 94.0–98.2) and 96.8% (95% CI 94.7–98.2) based on TTE at index presentation and within 30 and 180 days post discharge, respectively. Among patients with a positive AI-ECG screen, 60% (65/109) had an LVEF < 50% (Figure 1) and over half of these cases (37/65, 57%) had an LVEF ≤ 35%. Positive predictive values for LVEF ≤ 35% were 33.9% (95% CI 25.2–43.6) and 34.1% (95% CI 25.9–43.1) at 30 and 180 days post discharge, respectively.
Left ventricular ejection fraction at echocardiogram according to artificial intelligence–augmented electrocardiography screen results. Proportion of patients with left ventricular ejection fraction <35%, between 35 and 50%, and >50% among those with a positive and negative artificial intelligence–augmented electrocardiography screen and an echocardiogram available at the index event or within 30 days in the overall cohort.
Prognostic performance of artificial intelligence–augmented electrocardiography for left ventricular systolic dysfunction for major adverse cardiovascular event
Most patients (1913/1977, 97%) had post-discharge follow-up data with a median follow-up of 24.4 months. Post-discharge MACE occurred in 24% of the overall cohort. Figure 2 shows the continuous relationship between an increasing probability of LVSD as determined by the AI-ECG algorithm and the corresponding increasing HR for MACE and mortality in the overall cohort. AI-ECG scores as a continuous variable (log) were strong predictors of MACE (adjusted HR 1.40, 95% CI 1.34–1.47) and mortality (adjusted HR 1.34; 95% CI 1.25–1.42). AI-ECG abnormalities were not associated with MACE on adjusted analyses (adjusted HR 1.22, 95% CI 0.91–1.64).
Relationship between artificial intelligence–augmented electrocardiography probabilities at risk of major adverse cardiovascular event and mortality. Continuous relationship between a probability of left ventricular systolic dysfunction as determined by the artificial intelligence–augmented electrocardiography algorithm and the corresponding risk (hazard ratio) for major adverse cardiovascular event (A) and mortality (B) in the overall cohort.
Outcomes stratified according to AI-ECG screen results are shown in Table 2. Patients with a positive AI-ECG screen had a higher rate of MACE than those with a negative AI-ECG screen (48 vs. 21%, P < 0.0001; adjusted HR 1.39, 95% CI 1.11–1.75; Graphical Abstract, Figure 3). These outcomes were driven by a higher rate of deaths (32 vs. 14%, P < 0.0001; adjusted HR 1.26, 95% CI 0.95–1.66) and HF hospitalizations (26 vs. 6.1%, P < 0.0001; adjusted HR 1.75; 95% CI 1.25–2.450; Graphical abstract; Figure 3, Supplementary material online, Table S3). A separate multivariable model including adjudicated ECG abnormalities (ischaemic/non-ischaemic and sinus/non-sinus) demonstrated that AI-ECG remained a significant, independent predictor of MACE (adjusted HR 1.33, 95% CI 1.06–1.67). Sensitivity analyses excluding patients with a history of HF demonstrated similar MACE outcomes (26 vs. 15%, P = 0.002; adjusted HR 1.99, 95% CI 1.35–2.92).
Survival curves according to artificial intelligence–augmented electrocardiography screen results. Kaplan–Meier survival curves for the composite outcome of major adverse cardiovascular events (A) and death (B) according to artificial intelligence–augmented electrocardiography results in the overall cohort.
Event rates of major adverse cardiovascular events at 2 years, overall and by negative vs. positive artificial intelligence–augmented electrocardiography screen
| Total n = 1977 | Negative AI-ECG | Positive AI-ECG | P-value | |
|---|---|---|---|---|
| n = 1729 | n = 248 | |||
| Post-discharge MACE, n (%) | 478 (24) | 359 (21) | 119 (48) | <0.0001 |
| Death, n (%) | 325 (16) | 247 (14) | 78 (32) | <0.0001 |
| Acute MI, n (%) | 41 (2.1) | 34 (2.0) | 7 (2.8) | 0.38 |
| HF hospitalization, n (%) | 169 (8.5) | 105 (6.1) | 64 (26) | <0.001 |
| Stroke or TIA, n (%) | 69 (3.5) | 56 (3.2) | 13 (5.2) | 0.11 |
| New-onset AF, n (%) | 42 (2.1) | 37 (2.1) | 5 (2.0) | 0.90 |
| Total n = 1977 | Negative AI-ECG | Positive AI-ECG | P-value | |
|---|---|---|---|---|
| n = 1729 | n = 248 | |||
| Post-discharge MACE, n (%) | 478 (24) | 359 (21) | 119 (48) | <0.0001 |
| Death, n (%) | 325 (16) | 247 (14) | 78 (32) | <0.0001 |
| Acute MI, n (%) | 41 (2.1) | 34 (2.0) | 7 (2.8) | 0.38 |
| HF hospitalization, n (%) | 169 (8.5) | 105 (6.1) | 64 (26) | <0.001 |
| Stroke or TIA, n (%) | 69 (3.5) | 56 (3.2) | 13 (5.2) | 0.11 |
| New-onset AF, n (%) | 42 (2.1) | 37 (2.1) | 5 (2.0) | 0.90 |
AF, atrial fibrillation/flutter; HF, heart failure; hs-cTnT, high-sensitivity-cardiac troponin T; MI, myocardial infarction; URLs, upper-reference limits; TIA, transient ischaemic attack.
Event rates of major adverse cardiovascular events at 2 years, overall and by negative vs. positive artificial intelligence–augmented electrocardiography screen
| Total n = 1977 | Negative AI-ECG | Positive AI-ECG | P-value | |
|---|---|---|---|---|
| n = 1729 | n = 248 | |||
| Post-discharge MACE, n (%) | 478 (24) | 359 (21) | 119 (48) | <0.0001 |
| Death, n (%) | 325 (16) | 247 (14) | 78 (32) | <0.0001 |
| Acute MI, n (%) | 41 (2.1) | 34 (2.0) | 7 (2.8) | 0.38 |
| HF hospitalization, n (%) | 169 (8.5) | 105 (6.1) | 64 (26) | <0.001 |
| Stroke or TIA, n (%) | 69 (3.5) | 56 (3.2) | 13 (5.2) | 0.11 |
| New-onset AF, n (%) | 42 (2.1) | 37 (2.1) | 5 (2.0) | 0.90 |
| Total n = 1977 | Negative AI-ECG | Positive AI-ECG | P-value | |
|---|---|---|---|---|
| n = 1729 | n = 248 | |||
| Post-discharge MACE, n (%) | 478 (24) | 359 (21) | 119 (48) | <0.0001 |
| Death, n (%) | 325 (16) | 247 (14) | 78 (32) | <0.0001 |
| Acute MI, n (%) | 41 (2.1) | 34 (2.0) | 7 (2.8) | 0.38 |
| HF hospitalization, n (%) | 169 (8.5) | 105 (6.1) | 64 (26) | <0.001 |
| Stroke or TIA, n (%) | 69 (3.5) | 56 (3.2) | 13 (5.2) | 0.11 |
| New-onset AF, n (%) | 42 (2.1) | 37 (2.1) | 5 (2.0) | 0.90 |
AF, atrial fibrillation/flutter; HF, heart failure; hs-cTnT, high-sensitivity-cardiac troponin T; MI, myocardial infarction; URLs, upper-reference limits; TIA, transient ischaemic attack.
Analyses according to adjudicated diagnoses demonstrated that when compared with patients with a negative AI-ECG screen, those with a positive AI-ECG screen had higher rates of MACE in patients classified as having myocardial injury (65 vs. 42%, P < 0.001; adjusted HR 1.44, 95% CI 1.10–1.87), with no difference observed in Type 2 MI (61 vs. 47%, P = 0.29), and a trend towards higher MACE in Type 1 MI (33 vs. 15%, P = 0.096; see Supplementary material online, Tables S3 and S4).
The conjoint use of hs-cTnT and AI-ECG screen for 2 years risk stratification for MACE and mortality are shown in Graphical abstract and Supplementary material online, Table S5. Together, hs-cTnT and AI-ECG resulted in the following MACE rates and adjusted HRs: hs-cTnT < 99th percentile and negative AI-ECG: 116/1176 (11%) (reference), hs-cTnT < 99th percentile and positive AI-ECG: 28/107 (26%; adjusted HR 1.54, 95% CI 1.01–2.36), hs-cTnT > 99th percentile and negative AI-ECG: 233/553 (42%; adjusted HR 2.12, 95% CI 1.66–2.70), and hs-cTnT > 99th percentile and positive AI-ECG: 91/141 (65%; adjusted HR 2.83, 95% CI 2.06–3.87; Figure 4). Similar findings were observed for mortality: hs-cTnT < 99th percentile and negative AI-ECG: 82/1176 (7.0%) (reference), hs-cTnT < 99th percentile and positive AI-ECG: 18/107 (17%; adjusted HR 1.52, 95% CI 0.90–2.58), hs-cTnT > 99th percentile and negative AI-ECG: 165/553 (30%; adjusted HR 2.02, 95% CI 1.51–2.70), and hs-cTnT > 99th percentile and positive AI-ECG: 60/141 (43%; adjusted HR 2.36, 95% CI 1.60–3.48). A multivariable model assessing the interaction between hs-cTnT and AI-ECG showed no evidence of statistical interaction (P = 0.98).
Survival curves according to the combination of high-sensitivity-cardiac troponin T and artificial intelligence–augmented electrocardiography screen results. Kaplan–Meier survival curves for the composite outcome of major adverse cardiovascular events (A) and death (B) according to the combinations of high-sensitivity-cardiac troponin T and artificial intelligence–augmented electrocardiography results in the overall cohort.
Discussion
The present study is the first to evaluate the role of an AI-ECG algorithm for LVSD-risk stratification in ED patients undergoing hs-cTnT measurement for suspected ACS. Our study provides several unique findings. First, using a validated AI-ECG algorithm for LVSD,16–19,22 we identified that at least 1 in 10 patients undergoing hs-cTnT measurement in the ED are at risk for LVSD as identified using this AI-ECG algorithm. Second, patients with concentrations above the sex-specific 99th percentile URLs are more likely than those without hs-cTnT increases to have positive AI-ECG screens for LVSD. Third, based on diagnoses established using the Fourth UDMI, patients with Type 1 MI had the highest risk for LVSD as 38% of these cases had positive AI-ECG screens, followed by similar frequencies of 20% in patients with myocardial injury and Type 2 MI. Fourth, AI-ECG probabilities for LVSD have important prognostic implications. Specifically, the higher the AI-ECG score, the higher the risk for MACE, including mortality. Our study demonstrates that patients with positive AI-ECG screens for LVSD are at significantly higher risk for MACE when compared with those with negative AI-ECG. Last, the conjoint use of hs-cTnT and AI-ECG facilitates the risk stratification for MACE and mortality.
Our findings have important clinical implications. Increases in hs-cTn above the 99th percentile URLs are universally recognized as indicative of acute or chronic myocardial injury1 and can help identify patients with previously unrecognized heart disease. These patients likely benefit from additional cardiovascular testing, among which echocardiography for LVSD is one of the most important evaluations given the prognostic and therapeutic implications.12,13 Clinical practice guidelines recommend echocardiography in intermediate-risk patients with chest pain and indicate that the presence of LVSD in those with suspected ACS reclassifies patients as high risk.3 Barriers to obtaining timely echocardiography are frequent. Likewise, echocardiographic evaluations may increase length of stay because of the challenges in obtaining them in an expeditious manner, which can further contribute to overcrowding and increase costs. While echocardiography provides valuable information, it is likely not feasible nor cost-effective in all patients with myocardial injury, particularly with the transition to hs-cTn when increases above the 99th percentile can be more common.4 Indeed, different American and European studies have shown that echocardiography is often not performed in a substantial percentage of patients with myocardial injury and Type 2 MI,5,14,15 even though the adverse prognosis related to these conditions has been extensively elucidated.3 Therefore, clinicians will benefit from rapid and cost-effective methods to facilitate triaging and resource utilization. Our study demonstrates how AI-ECG for LVEF can facilitate the detection of patients at high risk for LVSD in whom echocardiography should be performed, when compared with those at low risk in whom evaluations can be avoided or deferred. Accordingly, the role of AI-ECG would be not to replace echocardiography, but to permit systematic LVSD assessment in all patients undergoing hs-cTn measurements and a standard 12-lead ECG, particularly those with elevated hs-cTn, to assess their risk and the indication and timing for further evaluations. The pragmatic applications of AI-ECG algorithms are evolving, with emerging evidence evaluating AI-ECG with a single-lead as a point-of-care screening for HF with reduced ejection fraction.23
The application of AI-ECG for LVSD in ED patients undergoing hs-cTnT measurement provides objective and rapidly available information to improve triage and help guide clinicians in deciding whether additional evaluations are necessary. Our findings are based on an AI-ECG algorithm that has been extensively validated16–19,22 for the prediction of LVSD (LVEF ≤ 35). Using paired 12-lead ECG and echocardiographic data from 44 959 patients, it was trained in a convolutional neural network and tested in an independent set of 52 870 patients where it resulted in an area under the curve (AUC) of 0.93.16 It was subsequently validated in a large prospective cohort17 where an AUC of 0.91 was observed in 6008 patients with a TTE within a year from ECG and in 3874 patients with a TTE within <1 month of ECG. The algorithm has been validated in different clinical settings, including in ED patients presenting with dyspnoea18 with a TTE available with an AUC of 0.89 and accuracy of 85.9%. Among critical patients in the cardiac intensive care unit,19 including 55% with an ACS, the AI-ECG had an AUC of 0.83 for discrimination of LVSD. It was also evaluated in the ECG AI-Guided Screening for Low Ejection Fraction randomized controlled trial, where 11 573 patients were evaluated in the intervention arm in which clinicians had access to AI-ECG results.22
Our data demonstrates that AI-ECG provides prognostic information beyond that offered by standard morphological ischaemic or rhythm abnormalities identified by standard ECG interpretation. Further, because HF is an important covariate, multivariable models included a history of HF and sensitivity analyses were also performed that excluded patients with a prior history of HF. Our data consistently showed across various analyses that AI-ECG remained an important predictor of MACE, including a higher risk of HF hospitalization.
Increasing AI-ECG values were associated with progressively increasing risk for MACE and mortality during follow-up. Patients with an abnormal AI-ECG had a higher risk for MACE irrespective of hs-cTnT results. When we analysed the 2-year outcomes using hs-cTnT and AI-ECG together for risk stratification, when compared with patients with hs-cTnT ≤99th percentile and negative AI-ECG, there was a higher risk for adverse outcomes in those with hs-cTnT ≤ 99th percentile and a positive AI-ECG screen, followed by those with hs-cTnT >99th percentile and a negative AI-ECG screen, with the highest risk for MACE and mortality observed in those with hs-cTnT >99th percentile and a positive AI-ECG screen in which 2-year MACE was 65% and mortality 43%. These results support the conjoint use of hs-cTn and AI-ECG. The clinical use of this AI-ECG algorithm together with hs-cTnT measurements in the ED setting could enhance risk stratification and facilitate triage, especially in those with elevated hs-cTn concentrations. Similar to our findings, Mahayni et al.24 recently investigated the prognostic performance of the same AI-ECG screen among patients with known LVEF > 35% undergoing valve or coronary bypass surgery and reported that patients with a positive AI-ECG screen were at higher risk of long-term mortality compared with those with a normal AI-ECG screen.
We also examined the diagnostic performance of AI-ECG in identifying ED patients with LVSD, to verify its consistency in this setting. A positive AI-ECG screen was helpful in identifying patients with a reduced LVEF, with 60% of those with a positive AI-ECG screen having an LVEF < 50% at TTE at index presentation or within 30 days post discharge. Even though not all patients with a positive AI-ECG screen have LVD (LVEF < 35%) and a poor positive predictive value was observed for patients with LVEF < 35%, it is important to emphasize that a positive AI-ECG screen identified a high proportion of patients with LVEF < 50%, which is an actionable threshold for therapeutic interventions given it reflects an abnormal EF25 that is associated with a worse prognosis. Most importantly, the AI-ECG algorithm is a robust tool for excluding LVSD, with 87% of those with a negative AI-ECG screen having a preserved LVEF ≥ 50%, and negative predictive values of ∼97% for LVEF ≤ 35% observed in those with available TTE.
There are some limitations to this study. First, this is a retrospective, observational study, with the limitations related to the study design. Second, diagnostic misclassification between myocardial injury and infarction is possible. Third, echocardiography was performed on clinical indication and therefore not available for every patient enrolled in the study, which can potentially introduce selection bias. We also did not collect echocardiographic data prior to the index encounter, so we could not perform additional sensitivity analysis excluding patients with a known reduced LVEF; however, we performed sensitivity analyses based on a history of heart failure. Likewise, while the primary aim of the present study was to evaluate the retrospective application of this previously developed and validated AI-ECG algorithm for LVSD, we do report diagnostic performance for LVSD as a secondary analysis among patients that underwent echocardiography on clinical indication. These analyses are secondary and based on a modest number of patients with available TTE and larger studies including systematic measurement of hs-cTnT, 12-lead ECG, and echocardiography would be needed to allow more sophisticated diagnostic performance analysis. Fourth, we recognize that echocardiography provides additional information beyond LVEF that may influence clinical decision-making. Last, given the sample size and number of events, larger studies with more events are needed to validate our findings, and ideally prospective studies to determine whether its implementation improves outcomes.
Conclusion
Among ED patients evaluated with hs-cTnT, at least 1 in 10 are at risk for LVSD using AI-ECG. These patients are at high risk of long-term MACE. The conjoint use of hs-cTnT and AI-ECG facilitates risk stratification. This innovative approach may help identify patients in which echocardiography should be obtained or deferred based on their individualized risk for LVSD, and such information facilitates risk stratification for adverse outcomes when used in isolation, as well when used together with hs-cTnT.
Supplementary material
Supplementary material is available at European Heart Journal: Acute Cardiovascular Care.
Acknowledgements
This publication was made possible in part by the Mayo Clinic CTSA through grant number UL1TR002377 from the National Center for Advancing Translational Sciences (NCATS), a component of the National Institutes of Health (NIH).
Funding
None declared.
Data availability
The data underlying this article are available in the article and in its online Supplementary Material.
References
Author notes
Conflict of interest: Y.S.: advisory board/speaker for Roche Diagnostics and advisory board Abbott Diagnostics. Patent #20210401347. A.S.J. has consulted or presently consults for most of the major diagnostics companies, including Beckman-Coulter, Abbott, Siemens, ET Healthcare, Ortho Diagnostics, Roche, Radiometer, Sphingotec, RCE Technologies, and Amgen and Novartis. All other authors have nothing to disclose.
Comments