Articles

Diagnostic accuracy of a machine learning algorithm using

point-of-care high-sensitivity cardiac troponin I for rapid
rule-out of myocardial infarction: a retrospective study
Betül Toprak*, Hugo Solleder*, Eleonora Di Carluccio, Jaimi H Greenslade, William A Parsonage, Karen Schulz, Louise Cullen, Fred S Apple,
Andreas Ziegler, Stefan Blankenberg, on behalf of the Artificial Intelligence in Suspected Myocardial Infarction Study (ARTEMIS) group†

Summary
Background Point-of-care (POC) high-sensitivity cardiac troponin (hs-cTn) assays have been shown to provide similar Lancet Digit Health 2024
analytical precision despite substantially shorter turnaround times compared with laboratory-based hs-cTn assays. We Published Online
applied the previously developed machine learning based personalised Artificial Intelligence in Suspected Myocardial August 29, 2024
https://doi.org/10.1016/
Infarction Study (ARTEMIS) algorithm, which can predict the individual probability of myocardial infarction, with a
S2589-7500(24)00191-2
single POC hs-cTn measurement, and compared its diagnostic performance with standard-of-care pathways for rapid
*Contributed equally
rule-out of myocardial infarction.
†A full list of collaborators is
provided at the end of this Article
Methods We retrospectively analysed pooled data from consecutive patients of two prospective observational cohorts in Department of Cardiology
geographically distinct regions (the Safe Emergency Department Discharge Rate cohort from the USA and the Suspected (B Toprak MD, Prof A Ziegler PhD,
Acute Myocardial Infarction in Emergency cohort from Australia) who presented to the emergency department with Prof S Blankenberg MD) and
suspected myocardial infarction. Patients with ST-segment elevation myocardial infarction were excluded. Safety and University Center of
Cardiovascular Science
efficacy of direct rule-out of myocardial infarction by the ARTEMIS algorithm (at a pre-specified probability threshold (B Toprak, Prof S Blankenberg)
of <0·5%) were compared with the European Society of Cardiology (ESC)-recommended and the American College of and Department for
Cardiology (ACC)-recommended 0 h pathways using a single POC high-sensitivity cardiac troponin I (hs-cTnI) Population Health Innovation
measurement (Siemens Atellica VTLi as investigational assay). The primary diagnostic outcome was an adjudicated (B Toprak, Prof S Blankenberg),
University Heart and Vascular
index diagnosis of type 1 or type 2 myocardial infarction according to the Fourth Universal Definition of Myocardial Center, University Medical
Infarction. The safety outcome was a composite of incident myocardial infarction and cardiovascular death (follow-up Center Hamburg-Eppendorf,
events) at 30 days. Additional analyses were performed for type I myocardial infarction only (secondary diagnostic Hamburg, Germany; German
outcome), and for each cohort separately. Subgroup analyses were performed for age (<65 years vs ≥65 years), sex, Center for Cardiovascular
Research (DZHK), Partner Sites
symptom onset (≤3 h vs >3 h), estimated glomerular filtration rate (<60 mL/min per 1·73 m² vs ≥60 mL/min per 1·73 m²), Hamburg/Kiel/Luebeck,
and absence or presence of arterial hypertension, diabetes, a history of coronary artery disease, myocardial infarction, or Hamburg, Germany (B Toprak,
heart failure, smoking, and ischaemic electrocardiogram signs. Prof A Ziegler,
Prof S Blankenberg); Cardio-
CARE, Medizincampus Davos,
Findings Among 2560 patients (1075 [42%] women, median age 58 years [IQR 48·0–69·0]), prevalence of myocardial Davos, Switzerland
infarction was 6·5% (166/2560). The ARTEMIS-POC algorithm classified 899 patients (35·1%) as suitable for rapid (H Solleder PhD,
rule-out with a negative predictive value of 99·96% (95% CI 99·64–99·96) and a sensitivity of 99·68% (97·21–99·70). E Di Carluccio MSc,
Prof A Ziegler); Emergency and
For type I myocardial infarction only, negative predictive value and sensitivity were both 100%. Proportions of missed
Trauma Centre, Royal Brisbane
index myocardial infarction (0·05% [0·04–0·42]) and follow-up events at 30 days (0·07% [95% CI 0·06–0·59]) were and Women’s Hospital,
low. While maintaining high safety, the ARTEMIS-POC algorithm identified more than twice as many patients as Brisbane, QLD, Australia
eligible for direct rule-out compared with guideline-recommended ESC 0 h (15·2%) and ACC 0 h (13·8%) pathways. (J H Greenslade PhD,
Prof L Cullen MD); Australian
Superior efficacy persisted across all clinically relevant subgroups.
Centre for Health Services
Innovation, School of Public
Interpretation The patient-tailored, medical decision support ARTEMIS-POC algorithm applied with a single POC Health and Social Work,
hs-cTnI measurement allows for very rapid, safe, and more efficient direct rule-out of myocardial infarction than Queensland University of
Technology, Brisbane, QLD,
guideline-recommended pathways. It has the potential to expedite the safe discharge of low-risk patients from the
Australia
emergency department including early presenters with symptom onset less than 3 h at the time of admission and (Prof W A Parsonage MD,
might open new opportunities for the triage of patients with suspected myocardial infarction even in ambulatory, Prof L Cullen); Department of
preclinical, or geographically isolated care settings. Laboratory Medicine and
Pathology, University of
Minnesota, Minneapolis, MN,
Funding The German Center for Cardiovascular Research (DZHK). USA (K Schulz DC,
Prof F S Apple PhD); Faculty of
Copyright © 2024 The Author(s). Published by Elsevier Ltd. This is an Open Access article under the CC BY Medicine, University of
Queensland, Brisbane, QLD,
4.0 license. Australia (Prof L Cullen);
Hennepin Healthcare Research
Introduction patients with indicative symptoms, only 5% to 25% are Institute, Minneapolis, MN,
Chest pain is the leading cause of presentation to the finally diagnosed with an acute myocardial infarction or USA (Prof F S Apple); School of
Mathematics, Statistics and
emergency department globally.1 However, among require immediate treatment.2 The identification of

www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2 1

 Articles

Computer Science, University

of KwaZulu-Natal, Research in context
Pietermaritzburg, South Africa
(Prof A Ziegler) Evidence before this study provides the first (to our knowledge) retrospective
Correspondence to: Myocardial infarction is a major global health issue. International application of a patient-tailored algorithm in conjunction
Prof Stefan Blankenberg, guidelines recommend standardised triage pathways to diagnose with a single POC high-sensitivity cardiac troponin I
Department of Cardiology, and rule-out myocardial infarction in emergency departments. measurement and its suggested threshold for rapid rule-out
University Heart and Vascular
These conventional pathways rely on fixed assay-specific high- of myocardial infarction in two geographically distinct
Center, University Medical Center
Hamburg-Eppendorf, sensitivity cardiac troponin (hs-cTn) concentrations and require cohorts. The ARTEMIS-POC algorithm yields doubled efficacy
20246 Hamburg, Germany central laboratory structures. In addition, they are not applicable with maintained high safety compared with guideline-
s.blankenberg@uke.de to patients with early symptom onset, and neglect clinical recommended pathways. Further, it performs consistently
variables. A global application including a personalised and well across clinically challenging subgroups, including
universal approach for clinical decision making is lacking. Machine patients with early symptom onset.
learning is increasingly used to address these limitations by
Implications of all the available evidence
combining distinct variables to achieve high diagnostic accuracy.
Our study builds upon previous work within the ARTEMIS
We systematically searched PubMed for studies published from
framework and extends the concept of machine learning for
database inception up to April 5, 2024, using the keywords:
rule-out of myocardial infarction to its use in conjunction
“machine learning”, “myocardial infarction”, “diagnosis”, and
with POC hs-cTn. The ARTEMIS-POC algorithm can be
“troponin” without any language restrictions. We identified three
calibrated and customised according to local care priorities
machine learning algorithms for suspected myocardial infarction
(eg, rural areas) by providing flexibility of the selection of
that were validated after derivation: myocardial-ischaemic-injury-
diagnostic thresholds. Further, it has the potential to reduce
index (MI³), Collaboration for the Diagnosis and Evaluation of
rule-out time to only a few minutes even in patients with
Acute Coronary Syndrome (CoDE-ACS), and Artificial Intelligence
early symptom onset, safely redirect low-risk patients away
in Suspected Myocardial Infarction Study (ARTEMIS). Among
from the emergency department, and allow for an extension
them, the personalised ARTEMIS algorithm that combines hs-cTn
of rule-out of suspected myocardial infarction from high-
values with routinely available clinical variables to estimate the
resource emergency departments with access to a central
individual probability of myocardial infarction is the first to be
laboratory 24 h a day towards alternative health-care
applicable with numerous available hs-cTn assays and to
settings, including preclinical or ambulatory care. The current
specifically allow for the integration of point-of-care (POC)
application of the ARTEMIS algorithm in existing study
hs-cTn. In contrast to other machine learning-based diagnostic
cohorts allows for its calibration to multiple clinical and local
algorithms for suspected myocardial infarction, the ARTEMIS
settings. To achieve real-world clinical implementation,
algorithm does not require measurement of further biomarkers
investigation of outcomes such as feasibility, sustainability,
and can be flexibly and universally applied with most available
and cost-effectiveness of the ARTEMIS algorithm is still
hs-cTn assays including POC tests.
warranted. Overall, the integration of POC hs-cTn into an easy
Added value of this study to use, calibrated algorithm for suspected myocardial
By combining the previously derived, validated, and infarction may allow for its uniform clinical application in any
generalised ARTEMIS algorithm with POC hs-cTn, our study setting.

individuals at low risk of myocardial infarction who will pathways do not allow for direct rule-out of myocardial
not benefit from hospital admission or who qualify for infarction in early presenters based on a single hs-cTn
early discharge has gained growing interest. measurement.
Together with the introduction of high-sensitivity To address these limitations, the Artificial Intelligence
cardiac troponin (hs-cTn) assays, a continuous evolution in Suspected Myocardial Infarction Study (ARTEMIS)
of diagnostic pathways towards rapid rule-out protocols algorithm was developed and externally validated in
has emerged to achieve this goal.3–5 Such clinical two German cohorts, including patients with symptoms
decision pathways (ie, as endorsed by the European suggestive of an acute myocardial infarction, and
Society of Cardiology [ESC]6 or the American College of most importantly, was shown to be generalisable in
Cardiology [ACC])7 rely on fixed assay-specific hs-cTn 13 international cohorts across four continents. This
thresholds and are limited by turnaround times of labo- personalised diagnostic algorithm estimates the individ-
ratory-based hs-cTn testing. Additionally, these clinical ual probability of myocardial infarction by integrating
decision pathways do not systematically account for routinely available clinical variables of interest, using
important clinical variables that are known to influence single or serial concentrations of six different hs-cTn
hs-cTn concentrations and pretest probability of myo- assays, and allowing for flexible timing of serial blood
cardial infarction, such as age, sex, symptom onset, testing.12 The algorithm is optimised for the identification
electrocardiographic findings, and cardiovascular risk of both non-ST-elevation myocardial infarction type 1
factors.8–11 In addition, both guideline-recommended (T1MI) and type 2 (T2MI).

2 www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2

 Articles

If this patient-tailored algorithm is used in connection Patients from both cohorts in whom the treating
with modern hs-cTn point-of-care (POC) assays fulfilling physician performed investigation for ruling out or
the requirements of high sensitivity based on whole ruling in an acute myocardial infarction upon clinical
blood assessment,13,14 it has the potential to improve indication were included in the analyses of this current
immediate and safe rule-out of patients not only in the study. Patients with a symptom onset of less than 3 h
emergency department, but also in ambulatory and pre- were considered early presenters.
clinical care including rural settings in which hospital
infrastructure is insufficiently available. Procedures
We therefore aimed to retrospectively apply the Cardiac troponin assays
ARTEMIS algorithm and the guideline-recommended In SEIGE, fresh ethylene diamine tetraacetic acid plasma
pathways in conjunction with a single POC hs-cTn level was sampled for hs-cTnI measurement using the
at presentation and compare their safety and efficacy for Architect hs-cTnI assay (Abbott, IL, USA). In SAMIE, the
immediate rule-out in two geographically distinct cohorts Access hs-cTnI assay (Beckman Coulter, Brea, CA, USA)
of patients with suspected myocardial infarction. was used in routine clinical practice. Hs-cTnI measure-
ments used in routine care (Abbott or Beckman) were
Methods considered for adjudication of diagnoses in both cohort
Study cohorts studies.
Individual-level data from two prospective observational In SEIGE, lithium-heparinised whole blood was used
cohort studies, the US-based Safe Emergency Department to concurrently measure Siemens POC Atellica VTLi
Discharge Rate (SEIGE; registered with ClinicalTrials.gov, hs-cTnI. In SAMIE, lithium-heparinised plasma samples
NCT04772157; registration date: Feb 26, 2021) and were taken at the same time as the samples for the
the Australian Suspected Acute Myocardial Infarction standard-of-care hs-cTnI assay and stored at –80°C. These
in Emergency (SAMIE; registered with ACTRN, samples were later thawed and tested in a single batch on
ACTRN12621000053820; date of registration: Jan 21, 2021) the POC Atellica VTLi platform at the end of July, 2021.
cohort, were jointly and retrospectively analysed. The The POC Atellica VTLi assay, which is the investigational
STARD checklist for this study is provided in the appendix assay in both cohorts, has a limit of detection of See Online for appendix
(p 10). Study protocols for each cohort were approved by 1·24 ng/L. The lower limits of quantification are 3·7 ng/L
the respective institutional review boards (IRBs) under for whole blood and 2·1 ng/L for plasma. Further assay-
numbers HHRI 20–4828 and LNR/2020/QRBW/65773. specific analytical information is provided in the
All participants provided written informed consent. Based appendix (p 2).
on the study-specific IRBs, the need for additional ethical
approval to conduct the current analyses was waived. Adjudication of index diagnoses
The SEIGE study includes consecutive patients aged In each cohort, adjudication of diagnoses was performed
21 years or older presenting to the emergency department independently by two local specialists (cardiologists,
in whom high-sensitivity cardiac troponin I (hs-cTnI) emergency medicine specialists, or both) in a blinded
measurements were obtained upon clinical indication fashion. For adjudication, all available medical records
between October, 2020 and January, 2021 at the Hennepin were reviewed, including 12-lead ECG, findings from
Healthcare/Hennepin County Medical Center echocardiography and angiography if available, hs-cTnI
(Minneapolis, MN, USA). Patients undergoing investiga- concentrations as measured by the clinically applied
tion for suspected myocardial infarction were included if assay, and other clinical data. The Fourth Universal
they had serial cTnI measurements at minimum at Definition of Myocardial Infarction was used for the
baseline and 2 h after presentation and at least one 12-lead diagnosis of myocardial infarction. Patients with myocar-
electrocardiogram (ECG) available. Patients were excluded dial infarction were further classified as having T1MI or
if they had ST-segment elevation myocardial infarction, T2MI; T2MI required objective evidence or documenta-
were pregnant, declined to participate, or were unable to tion of oxygen supply–demand imbalance as the
provide informed consent. underlying pathogenesis of myocardial ischaemia.15,16
The SAMIE study includes consecutive patients aged
18 years or older who presented to one of five Australian Follow-up data at 30 days
hospitals between November, 2020 and September, 2021. Follow-up information (ie, incident myocardial infarc-
Blood samples for hs-cTn measurements were collected tion excluding the index myocardial infarction event, and
at presentation and 2–3 h later in all patients per protocol. cardiovascular death) was collected at 30 days after index
Patients were excluded if they had ST-segment elevation presentation to the emergency department using medical
myocardial infarction, were pregnant, transferred from and social security records.
another hospital, previously enrolled within 30 days, or
were unable or unwilling to provide informed consent. ARTEMIS diagnostic algorithm
Further details about both cohorts are provided in the The summary concept with all steps related to derivation,
appendix (p 2). validation, and generalisation of the ARTEMIS algorithm

www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2 3

 Articles

has been previously described.12 Details are provided in probability as a number between 0% and 100%. A myo-
the appendix (pp 3–6). In brief, feature selection for the cardial infarction probability threshold of less than 0·5%
ARTEMIS models was initially performed using different was previously proposed as the applicable direct rule-out
learning machines and 10-fold cross-validation across all threshold.12 Relevant information regarding all modelling
available hs-cTn assays and a total of 18 available steps is provided in an executive summary in the
candidate features. Eight clinical variables, including appendix (p 4). No modifications were made to the model
age, sex, symptom onset greater than 3 h, ischaemic in the current study.
signs in the ECG in line with guideline-recommended Among all available hs-cTn assays, the current analyses
definitions7 (details are provided in the appendix [p 4]), focus on the application of the POC VTLi hs-cTnI assay
heart rate, smoking status (including current and former in conjunction with the single hs-cTn measure-
smoking), hyperlipoproteinaemia, and family history of ment-based ARTEMIS algorithm and the pre-specified
coronary artery disease were kept after variable selection probability threshold.
For more on IFCC C-CB see in the single hs-cTn ARTEMIS model. This model
https://ifcc.org/ estimates the individual myocardial infarction Guideline-recommended pathways
According to the ESC 0 h pathway, immediate rule-out
was allowed if symptom onset was greater than 3 h and
All (N=2560) No index MI (n=2394) Index MI (n=166)
hs-cTnI was very low. A rapid rule-out threshold of less
Age, years 58·0 (48·0–69·0) 58·0 (47·0–68·0) 61·5 (53·0–74·5) than 4 ng/L was previously derived and validated for the
Sex POC VTLi assay.14 This threshold was used to apply
Male 1485 (58·0%) 1379 (57·6%) 106 (63·9%) the ESC 0 h pathway in our present study. According to
Female 1075 (42·0%) 1015 (42·4%) 60 (36·1%) the ACC 0 h pathway (ie, the modified ESC 0 h pathway),
Vital parameters immediate rule-out was possible with the presence of a
Heart rate, bpm 79·0 (68·0–94·0) 79·0 (68·0–93·0) 85·0 (72·0–104·0) non-ischaemic ECG, symptom onset 3 h or greater, and
Systolic blood pressure, mm Hg 135·0 (121·0–152·0) 134·0 (121·0–151·0) 141·0 (123·0–157·8) hs-cTnI below the respective limit of quantification.
Cardiac history Following the recommendations of the International
History of CAD 608 (23·8%) 544 (22·7%) 64 (38·6%) Federation of Clinical Chemistry and Laboratory
History of MI 429 (16·8%) 378 (15·8%) 51 (30·7%) Medicine–Committee of Clinical Applications of Cardiac
History of HF* 197 (7·7%) 172 (7·2%) 25 (15·1%) Bio-Markers (IFCC C-CB), hs-cTnI values were rounded
Risk factors to the nearest integer before applying the pathway rules.
Hypertension 1416 (55·3%) 1304 (54·5%) 112 (67·5%) Further details are provided in the appendix (p 7).
Hyperlipoproteinaemia 1201 (46·9%) 1098 (45·9%) 103 (62·0%)
Smoker (current or previous) 978 (38·2%) 908 (37·9%) 70 (42·2%) Outcomes
Family history of CAD 899 (35·1%) 850 (35·5%) 49 (29·5%) The primary diagnostic outcome was defined as the
Diabetes 673 (26·3%) 611 (25·5%) 62 (37·3%) adjudicated final diagnosis of non-ST-elevation myocar-
Index presentation data dial infarction (T1MI or T2MI) at index presentation to
Symptom onset >3 h 1243 (48·6%) 1156 (48·3%) 87 (52·4%)
the emergency department. The secondary diagnostic
Ischaemic ECG signs 39 (1·5%) 28 (1·2%) 11 (6·6%)
outcome comprised the adjudicated diagnosis of T1MI
Atellica VTLi POC hs-cTnI at 5·0 (3·1–10·3) 4·6 (3·0–8·9) 36·7 (16·4–118·8)
only. The safety outcome was a composite of incident
baseline, ng/L myocardial infarction and cardiovascular death at
Abbott hs-cTnI at baseline, ng/L† 8·0 (4·0–21·0) 7·0 (4·0–18·0) 59·5 (21·8–171·5) 30 days, excluding events that occurred during index
Beckman hs-cTnI at baseline, 3·0 (2·3–6·0) 3·0 (2·3–6·0) 61·0 (22·8–193·5) presentation.
ng/L‡
Type 1 MI§ 74 (2·9%) ·· 74 (44·6%) Statistical analysis
Type 2 MI§ 92 (3·6%) ·· 92 ( 55·4%) Baseline characteristics as provided in patients’ medical
ARTEMIS MI probability at 1·7 (0·1–7·2) 1·5 (0·1–6·6) 22·1 (7·4–49·8) records are displayed as absolute numbers (percentages)
baseline for categorical variables and median (quartiles) for con-
30-day outcomes tinuous variables. Patients with symptoms suggestive of
30-day MI 12 (0·5%) 10 (0·4%) 2 (1·2%) an acute myocardial infarction were enrolled in both
30-day cardiovascular death 21 (0·8%) 17 (0·7%) 4 (2·4%) registries making it imperative for clinicians to either
30-day composite¶ 33 (1·3%%) 27 (1·1%) 6 (3·6%) rule in or rule out a non-ST-elevation myocardial infarc-
Data are n (%) for categorical variables and median (IQR) for continuous variables and are provided for the pooled tion. Therefore, once the patient had agreed to participate
dataset (SEIGE and SAMIE cohorts) according to absence or presence of an adjudicated index diagnosis of myocardial in the study, all relevant data were collected. If data were
infarction (type 1 or type 2). CAD=coronary artery disease. ECG=electrocardiogram. HF=heart failure. hs-cTnl=high- missing, this was accidental. We thus assume the data to
sensitivity cardiac troponin I. MI=myocardial infarction. POC=point-of-care. *Known heart failure of any type. †Assay
used for adjudication in SEIGE. ‡Assay used for adjudication in SAMIE. §According to the Fourth Universal Definition of be missing completely at random and the missing at
MI. ¶Composite of MI and cardiovascular death at 30 days, excluding index events. random assumption to be implicitly fulfilled. Multiple
imputation of missing data (including ECG data in
Table 1: Baseline characteristics and 30-day outcomes in pooled cohorts
SAMIE) was performed by using 12 generalisation

4 www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2

 Articles

<0·1% <0·2% <0·3% <0·4% <0·5% <0·6% <0·7% <0·8% <0·9% <1%
NPV 99·96 99·96 99·96 99·96 99·96 99·96 99·75 99·74 99·73 99·71
(99·67–99·97) (99·66–99·96) (99·67–99·97) (99·65–99·96) (99·64–99·96) (99·63–99·96) (99·64–99·85) (99·66–99·80) (99·69–99·76) (99·69–99·73)
Sensitivity 99·80 99·75 99·74 99·70 99·68 99·65 98·61 98·45 98·30 98·14
(98·24–99·81) (97·85–99·77) (97·74–99·76) (97·40–99·73) (97·21–99·70) (96·99–99·68) (98·00–99·11) (97·44–99·22) (96·84–99·32) (96·20–99·41)
Rule-out 653 (25·5%) 748 (29·2%) 809 (31·6%) 854 (33·4%) 899 (35·1%) 939 (36·7%) 964 (37·7%) 988 (38·6%) 1018 (39·8%) 1046 (40·9%)
Index MI 0·04 0·04 0·04 0·05 0·05 0·05 0·24 0·26 0·27 0·29
(0·04–0·36) (0·04–0·39) (0·04–0·38) (0·04–0·41) (0·04–0·42) (0·04–0·43) (0·23–0·26) (0·22–0·30) (0·20–0·36) (0·17–0·43)
Follow-up event 0·05 0·06 0·06 0·06 0·07 0·08 0·09 0·10 0·10 0·11
(0·04–0·41) (0·05–0·51) (0·05–0·52) (0·06–0·54) (0·06–0·59) (0·07–0·71) (0·08–0·81) (0·09–0·88) (0·10–0·91) (0·10–0·95)
All effect estimates were pooled across imputed datasets for the jointly analysed cohorts (2560 patients). Data are median percent (95% CI) or n (%). Performance measures for direct rule-out by ARTEMIS-POC
algorithm are shown for a range of MI probabilities from <0·1% to <1%. MI=myocardial infarction. NPV=negative predictive value. POC=point-of-care.

Table 2: NPVs, sensitivities, and rule-out proportions for index MI, missed index MI, and follow-up events using the ARTEMIS-POC algorithm at different MI probability thresholds

cohorts within the ARTEMIS project. Details on these A

cohorts and the imputation procedure are provided in 100
the appendix (pp 8–9). Following multiple imputation,
calibration was performed across all imputed datasets.
Estimates were pooled with Rubin’s rule; details are
again provided in the appendix (p 9).
Percentage

To assess the diagnostic performance of the ARTEMIS- 95

POC algorithm across the spectrum of probability
thresholds for myocardial infarction, probabilities from
0·1% to 1% were considered in steps of 0·1%. Performance
measures for rapid rule-out were negative predictive value NPV
90
(NPV), sensitivity, and efficacy (ie, the proportion of Sensitivity
patients being ruled out), the proportion of missed index 0

myocardial infarction events, and the 30-day incidence of

B
myocardial infarction or cardiovascular death among 100
ruled-out individuals. Corresponding 95% CIs were
90
estimated to assess variability. Tests for statistical signifi-
cance were not applicable to our analyses, especially for the 80

comparison of the performances of the ARTEMIS 70

Proportion ruled out (%)

algorithm, ESC, and ACC pathways, where we did not 60

perform such tests due to the low number of events.
50
Additional analyses were performed for T1MI only (the
40
secondary diagnostic outcome), and per cohort (ie, for
SEIGE and SAMIE separately). 30
Subgroup analyses were performed on non-imputed 20
data for age (<65 years vs ≥65 years), sex, symptom onset 10
(≤3 h vs >3 h), estimated glomerular filtration rate
0
(<60 mL/min per 1·73 m² vs ≥60 mL/min per 1·73 m²), <0·1 <0·2 <0·3 <0·4 <0·5 <0·6 <0·7 <0·8 <0·9 <1·0
and absence or presence of arterial hypertension, diabetes, MI probability (%)
a history of coronary artery disease, myocardial infarction,
Figure 1: NPVs, sensitivities, and rule-out proportions for index MI using the ARTEMIS-POC algorithm
or heart failure, smoking, and ischaemic ECG signs. MI=myocardial infarction. NPV=negative predictive value.
Statistical analyses were performed in R version 4.2.0.
ethnicity were available in both cohorts. An acute myo-
Role of the funding source cardial infarction at index presentation was diagnosed in
Neither the funder nor any external party had a role in 166 (6·5%) of 2560 patients, of whom 74 patients (2·9%)
the data collection, analysis interpretation, writing of the had T1MI and 92 (3·6%) had T2MI. Individuals with an
manuscript, or the decision to submit for publication. adjudicated final diagnosis of myocardial infarction were
more likely to have a history of coronary artery disease,
Results myocardial infarction, heart failure, and traditional
Overall, 2560 patients were included from both cohorts. cardiovascular risk factors. A total of 1243 (48·6%) of
The median age was 58 years (IQR 48·0–69·0) and 2560 individuals had symptom onset greater than 3 h.
1075 (42·0%) were women (table 1). No data on race or Median Atellica VTLi POC hs-cTnI concentration was

www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2 5

 Articles

5·0 ng/L (IQR 3·1–10·3; non-myocardial infarction: 4·6 2560 individuals. The corresponding NPV and sensitivity
ng/L [3·0–8·9]; myocardial infarction: 36·7 ng/L were 100%. Similarly, the ACC 0 h pathway assigned
[16·4–118·8]). Cohort-specific baseline characteristics are 352 (13·8%) of 2560 individuals to rapid rule-out with
displayed in the appendix (p 12). both NPV and sensitivity reaching 100%. In comparison,
Among all 2560 patients, 33 (1·3%) experienced either the ARTEMIS-algorithm identified more than twice as
an acute myocardial infarction, cardiovascular death, or many individuals as eligible for rapid rule-out at the pre-
both within 30 days after index presentation (table 1). specified threshold of less than 0·5% while maintaining
Per-cohort follow-up data are shown in the appendix comparable safety (NPV and sensitivity; table 3).
(p 13). Among the patients of whom the ARTEMIS-POC
In pooled analyses, the ARTEMIS-POC algorithm algorithm missed an index myocardial infarction, none
maintained its direct rule-out performance across a had a T1MI event. No index myocardial infarction event
broad range of probability thresholds for myocardial was misclassified by either guideline-recommended
infarction (table 2, figure 1). Respective per-cohort pathways. At the same time, 30-day event rates were
analyses are provided in the appendix (pp 19–20). comparable between the ARTEMIS-POC algorithm, and
At a pre-specified probability threshold of less than the ESC and ACC 0 h pathways (table 3). The head-to-
0·5%, the ARTEMIS-POC algorithm classified head comparison of performance measures as compared
899 patients (35·1%) of 2560 as suitable for immediate with both guideline-recommended pathways for T1MI
rule-out. The corresponding NPV was 99·96% (95% CI only is provided in the appendix (p 14). Cohort-specific
99·64–99·96), and the sensitivity was 99·68% performance measures are displayed in the appendix
(97·21–99·70). For T1MI only, both NPV and sensitivity (p 15).
of direct rule-out reached 100% with the ARTEMIS-POC In all clinically relevant subgroups, the ARTEMIS-POC
algorithm (appendix p 14). Cohort-specific performances algorithm outperformed both guideline-recommended
are further displayed in the appendix (pp 14–15). pathways regarding rule-out proportions while preserv-
Across all datasets in pooled cohorts, the rate of missed ing high safety (figure 2, appendix pp 24–27). In patients
index myocardial infarction events (all patients with with symptom onset greater than 3 h, the ARTEMIS-
T2MI) was 0·05% (95% CI 0·04–0·42) among patients POC algorithm allowed for safe immediate rule-out of
being ruled out based on a probability threshold of less myocardial infarction in 591 (47·5%) of 1243 patients,
than 0·5%. Myocardial infarction, cardiovascular death, compared with 329 (26·5%) of 1243 patients with the
or both at 30 days after index presentation occurred in ESC 0 h and 294 (23·7%) of 1243 patients with the ACC 0
0·07% (0·06–0·59) of patients assigned to direct rule-out h pathways (appendix pp 26–27). Even in early presenters
(table 3). Details about missed events at index presenta- with symptom onset less than 3 h, to whom guide-
tion and within 30 days thereafter are provided in the line-recommended pathways are not applicable, up to
appendix (p 16). The ARTEMIS-POC algorithm used in one fifth (153/806; 19·0%) were eligible for immediate
conjunction with a probability threshold of less than rule-out based on a single POC measurement.
0·5% performed consistently well across distinct clini-
cally relevant subgroups (figure 2). Discussion
Rapid rule-out performance measures across a range of In two geographically distinct cohorts from the USA and
single POC VTLi cutoff concentrations, and using the Australia, we clinically applied a medical decision
same range for the ESC 0 h and ACC 0 h pathways, are support, machine learning-based algorithm combining
displayed in the appendix (pp 21–23). The application of clinical variables with a single POC hs-cTnI measure-
the ESC 0 h pathway allowed for an immediate rule-out ment to calculate the personalised risk of having an acute
of an acute myocardial infarction in 390 (15·2%) of myocardial infarction. We specifically validated its pre-
specified risk threshold for rapid rule-out of myocardial
infarction and compared its performance with guideline-
ARTEMIS-POC algorithm ESC 0 h* pathway ACC 0 h† pathway
(MI probability <0·5%)
recommended pathways in the emergency department.
The ARTEMIS-POC algorithm achieved high diagnostic
Rule-out 899 (35·1%) 390 (15·2%) 352 (13·8%)
accuracy and safety of immediate rule-out of myocardial
NPV 99·96 (99·64–99·96) 100 (100–100) 100 (100–100)
infarction, yielded more than twice the efficacy compared
Sensitivity 99·68 (97·21–99·70) 100 (100–100) 100 (100–100)
with guideline-endorsed pathways, and maintained its
Missed index MI 0·05 (0·04–0·42) 0 (0–0) 0 (0–0)
robust performance across all clinically relevant
Missed follow-up events 0·07 (0·06–0·59) 0·08 (0·07–0·71) 0·08 (0·07–0·70)
subgroups, including early presenters.
All effect estimates were pooled across the imputed datasets and include T1MI and T2MI. Data are n (%) or median The ARTEMIS algorithm allows for the estimation of
(95% CI). ACC=American College of Cardiology. ESC=European Society of Cardiology. MI=myocardial infarction.
an individual risk probability of myocardial infarction.
NPV=negative predictive value. POC=point-of-care. T1MI=type 1 myocardial infarction. T2MI=type 2 myocardial
infarction. *ESC 0 h pathway; POC cut-off <4 ng/L if symptom onset >3 h. †ACC 0 h pathway; POC cut-off <4 ng/L if It is tailored to patient-specific features, customised for
non-ischaemic electrocardiogram and symptom onset ≥3 h. use with multiple available hs-cTn assays, and does not
require any specific hs-cTn cutoffs and timepoints
Table 3: Diagnostic accuracies for index MI
of serial hs-cTn testing.12 Previously, it successfully

6 www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2

 Articles

Subgroup size (n) Sensitivity (95% CI) NPV (95% CI) Proportion ruled out
(95% CI)

Age (years)
<65 1718 99·28 (93·82–99·34) 99·95 (99·58–99·96) 43·39 (31·25–55·94)
≥65 842 100 (100–100) 100 (100–100) 18·48 (12·02–25·97)
Sex
Female 1072 99·99 (99·91–99·99) 100 (99·99–100) 42·42 (31·01–54·25)
Male 1485 99·53 (95·94–99·57) 99·92 (99·27–99·92) 29·97 (19·65–41·44)
Hypertension
No 1134 99·07 (92·02–99·14) 99·94 (99·46–99·94) 49·06 (40·97–57·18)
Yes 1416 100 (100–100) 100 (100–100) 24·75 (17·42–32·89)
Diabetes
No 1886 99·52 (95·82–99·56) 99·95 (99·61–99·96) 40·00 (30·08–50·36)
Yes 673 100 (100–100) 100 (100–100) 21·02 (18·21–23·97)
History of CAD
No 1934 99·39 (94·71–99·44) 99·95 (99·58–99·96) 40·65 (27·78–54·21)
Yes 608 100 (100–100) 100 (100–100) 17·79 (8·88–28·95)
History of MI
No 2115 99·43 (95·07–99·47) 99·95 (99·60–99·96) 38·59 (25·04–53·11)
Yes 429 100 (100–100) 100 (100–100) 18·20 (10·24–27·83)
History of heart failure
No 880 100 (100–100) 100 (100–100) 32·68 (29·96–35·45)
Yes 197 100 (100–100) 100 (100–100) 14·21 (11·56–17·09)
Smoking
No 1577 99·62 (96·70–99·65) 99·95 (99·56–99·95) 37·63 (29·39–46·24)
Yes 978 100 (100–100) 100 (100–100) 32·68 (20·77–45·86)
Symptom onset (h)
>3 1243 99·99 (99·92–99·99) 100 (99·99–100) 47·58 (20·58–75·37)
≤3 806 99·20 (93·10–99·26) 99·74 (97·73–99·76) 18·94 (16·65–21·34)
Ischaemic ECG signs
No 1038 100 (100–100) 100 (100–100) 29·73 (27·40–32·11)
Yes 39 100 (100–100) 100 (100–100) 17·95 (10·88–26·33)
eGFR (mL/min)
<60 594 99·99 (99·89–99·99) 99·99 (99·93–99·99) 13·72 (12·14–15·39)
≥60 1931 99·56 (96·23–99·60) 99·96 (99·62–99·96) 41·73 (36·05–47·51)

90 92 94 96 98 100

NPV

Figure 2: Sensitivities, NPVs, and rule-out proportions using the ARTEMIS-POC algorithm (MI probability <0·5%) across patient subgroups
CAD=coronary artery disease. ECG=electrocardiogram. eGFR=estimated glomerular filtration rate. MI=myocardial infarction. NPV=negative predictive value.
POC=point-of-care.

underwent derivation and validation based on acute myocardial infarction. This ability to safely tailor
two German cohorts of patients with suspected myo- medical decisions in a considerable magnitude of
cardial infarction, followed by its generalisation in patients in different clinical and regional settings might
13 international cohort studies. It is now applied with improve health care globally.17
its pre-specified cutoffs12 and different hs-cTn assays to The ARTEMIS-POC algorithm provides the opportu-
various external cohorts. nity to extend safe and rapid triage to pre-hospital settings
Here, we presented the retrospective application of the with general practitioners, cardiologists in primary care,
algorithm and its suggested diagnostic threshold to and ambulances. It might also be applied in settings
observational cohorts from the USA and Australia for without continuous access to laboratory-based testing,
rapid rule-out of myocardial infarction in the context of a such as rural areas. The flexibility of selecting diagnostic
POC hs-cTn system. By integrating a POC hs-cTn assay thresholds (according to local priorities of different
which has been proven to deliver accurate results within health-care providers) when using algorithms that
a few minutes,13 the algorithm further adds to its potential provide continuous risk scores and probabilities for
for the rapid triage of suspected myocardial infarction. myocardial infarction harbours a clear advantage over
Applying the ARTEMIS-POC algorithm on a single blood guideline-recommended pathways that use fixed hs-cTn
draw can potentially reduce door-to-rule-out time in the thresholds.12
emergency department in more than one third of The ARTEMIS-POC algorithm, including distinct
patients being admitted with symptoms suggestive of an patient-specific features, performed consistently across

www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2 7

 Articles

subgroups. Importantly, in early presenters to whom evaluation of clinical application and associated health-
recommended pathways are not applicable and second care costs will allow for adequate cost-benefit analysis.
hs-cTn testing is generally advised by guidelines6 (or even In general, the development and approval of medical
the additional measurement of biomarkers such as decision support tools require a rigorous stepwise
copeptin which is currently being discussed in the approach to become globally applicable. First, the develop-
context of accelerated rule-out18), approximately one fifth ment of a diagnostic algorithm should capitalise on robust
of patients would potentially be eligible for early approaches for integrating and weighting multiple
discharge based on the ARTEMIS-POC algorithm. If variables (ie, by machine learning to achieve high accuracy
nationwide approximately 7 million patients with [derivation phase]). Second, its performance needs to be
suspicion of myocardial infarction are evaluated in the validated in independent cohorts with individuals being
emergency department each year,19 the ARTEMIS-POC evaluated for the investigated clinical condition (validation
algorithm would (excluding approximately 750 000 annual phase). Third, as different pre-test probabilities likely exist
cases of ST-segment elevation myocardial infarction20) for the respective condition of interest across distinct
potentially allow for direct rule-out of myocardial infarc- regions, the diagnostic model needs to be applied in
tion in more than 1·25 million patients, even including a various clinical settings with methods allowing for its
considerable proportion of patients who present early adaptation to local settings (ie, calibration [generalisation
after symptom onset. phase]). Fourth, pre-specified diagnostic thresholds
Previous work has already introduced machine should be applied to different cohorts to compare the
learning concepts for the evaluation of suspected algorithm’s application with the current standard-of-care
myocardial infarction (for comparison, see appendix pathways (application phase). Fifth, the algorithm and its
pp 17–18);21,22 the myocardial ischaemic injury index (MI³) diagnostic cutoffs need to be tested in prospective ran-
was the first among them, and relies on two predefined domised controlled trials to demonstrate clinical utility
variables, age and sex, and is restricted to one specific (clinical trial phase). Lastly, after the regulatory processes
hs-cTnI assay. MI3 does not allow for immediate rule-out such as certification have been completed, real-world
of low-risk patients as it requests serial troponin testing, clinical implementation requires investigating outcomes,
and demonstrated rather poor calibration.21 The recently including feasibility, sustainability, and cost-effectiveness
published Collaboration for the Diagnosis and Evaluation (implementation phase; appendix p 28).
of Acute Coronary Syndrome (CoDE-ACS) score was The ARTEMIS models have not yet been prospectively
derived and validated in a large international patient tested as part of routine clinical practice. To exploit its
population,22 and compared with three guideline-recom- full use case potential in different settings including
mended pathways.23 Beyond its applicability with only in-hospital, ambulatory, and preclinical care, randomised
one hs-cTnI central laboratory-based assay, it requires the controlled clinical trials evaluating the implementation
addition of further biomarkers to calculate a risk score. of the ARTEMIS algorithm as a medical decision support
Furthermore, the CoDE-ACS score was designed to device are warranted. We are currently in the process of
mainly identify T1MI.22 Another emerging concept is obtaining certification for the ARTEMIS algorithm as a
artificial intelligence-powered ECG analysis for detecting medical device, as well as developing web-based software
occlusion myocardial infarction, which is an easy-to-ap- for clinical use of the algorithm.
ply method but needs prospective validation.24 As a future Some limitations should be considered. First, although
perspective, incorporating digitised ECG data into risk the ARTEMIS-POC algorithm is potentially applicable to
models might further improve triage of suspected myo- primary care settings with lower pre-test probability and
cardial infarction. prevalence of myocardial infarction than in the
In contrast to guideline-based pathways, applying such emergency department, recalibration will enable for its
algorithms in clinical practice presupposes regulatory applicability in these settings as the algorithm was
approval, digital implementation (eg, the development of derived and validated in emergency patients clinically
a native app or web-based interface), and devices for the assessed for suspected myocardial infarction.
algorithm to be run on, which is associated with costs Furthermore, the derivation and validation of ARTEMIS
and most likely translates into commercial purchase. In models were performed in German cohorts. Therefore,
the case of integrated biomarker measurements, costs applying these algorithms to other countries and global
for measurement, including sample preparation, regions is work in progress by using calibration methods
purchase and maintenance of systems, consumables, and performing prospective randomised clinical trials.
and quality control, have to be considered. In general, Although global generalisation was performed, data on
costs related to POC assays are usually higher than ethnicity were not available for the derivation, validation,
central laboratory-based testing and it remains unclear or generalisation cohorts, nor for the cohorts included in
whether reduced turnaround times really translate into the current study. It remains unclear how ethnicity might
reduced length of stay in the emergency department and impact performance. Second, the current work focused
health-care costs, and how patient flow will be affected if on rapid rule-out paths. We believe that the ability of
used in pre-hospital settings.13 Only prospective POC assays to provide reliable hs-cTn results within a

8 www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2

 Articles

few minutes is of considerable value in the context of Ehsan Mahmoodi, Siegfried Perez, Isuru Ranasinghe, Andrew Staib,
immediate rule-out of myocardial infarction, which Anna Zournazi, and Martin Than.
could translate into a rapid discharge of low-risk patients Contributors
from the emergency department without the need for BT, AZ, and SB conceived the idea for the study, interpreted the data, and
wrote the paper with input from all listed authors. HS, EDC, and
serial sampling. However, further investigations will AZ analysed the data. JHG, WAP, KS, LC, and FSA provided the data and
follow for the ARTEMIS-POC algorithm to demonstrate were involved in data interpretation and discussion of the study results.
its full potential, also for the triage of patients deemed All authors provided feedback and contributed significantly to the
intermediate-risk or high-risk. Analogous analyses are conducted research, analyses, and manuscript. BT and HS contributed
equally to this work. All authors were permitted to access the raw
also planned for each of the central laboratory-based datasets. BT, HS, EDC, AZ (lead statistician), and SB (corresponding
hs-cTn assays which are applicable with the ARTEMIS author) had full direct access to the underlying data and verified them.
algorithm. Third, it is important to note that the All authors have seen and approved the final version of the manuscript
ARTEMIS models were validated to estimate the risk of and accept the responsibility to submit the manuscript for publication.
an acute myocardial infarction in case of clinical Declaration of interests
suspicion, which does not include other conditions that BT is supported by a grant from the German Foundation of Heart
Research and Werner Otto Stiftung; is the co-recipient of a grant from
might cause similar symptoms, such as pulmonary the Ernst und Berta Grimmke-Stiftung (unrelated to this project); and
embolism or heart failure. The estimated probabilities of has received travel grants from the German Center for Cardiovascular
myocardial infarction must always be considered in con- Research (DZHK). BT is the local representative of the partner sites
junction with clinical assessment and should not be used Hamburg/Kiel/Luebeck in the Use and Access Committee of the
DZHK Heart Bank. JHG reports fellowship funding from Advance
as the only basis for decision making. Fourth, central Queensland and payment by the Emergency Medicine Foundation to her
laboratory-based hs-cTnI assays that were part of routine institution. WAP reports that his institution receives consulting fees
clinical care at the time of enrolment were used for from Siemens Healthineers. WAP reports acitivities at the Board
(executive and horonary treasurer) of the Cardiac Society of Australia and
real-time clinical management of patients and post-hoc
New Zealand. LC reports non-salaried grant support and provision of
adjudication of diagnoses in both cohorts, and not the assays through her institution from Abbott Diagnostics, Abbott Point-of-
investigational POC VTLi hs-cTnI assay. Fifth, POC Care (POC), Radiometer Pacific, Roche Diagnostics, Siemens
testing was performed by laboratory staff rather than Healthineers, and Beckman Coulter. LC also reports consulting fees
from Siemens Healthineers, Beckman Coulter, Quidel/Ortho, and
nursing staff. If applied to routine practice in the
payment or honoraria from Abbott Diagnostics, Roche Diagnostics, and
emergency department where nursing staff are in charge Glycardial. Both WAP and LC report research funding and provision of
of real-time bedside testing, aspects including education assays by Siemens Healthineers. FSA reports doing advisory board or
of the personnel, technical maintenance of the POC consultant activities for Mindray, Wefen, and Abbott Vascular, and non-
salaried grant support through the Hennepin Healthcare Research
device, and quality control measurements on a regular
Institute from Abbott Diagnostics, Abbott POC, Roche Diagnostics,
basis need to be considered and integrated into the Siemens Healthineers, Beckman Coulter, Quidel/Ortho, Becton
workflow.25 Sixth, ECG data that are part of the ARTEMIS Dickinson, and Sysmex, and fees for serving as Associate Editor for
algorithm were missing in the SAMIE cohort. To account Clinical Chemistry. AZ is a Board member of the Basel Biometric
Society. AZ and SB are listed as co-inventors of an international patent
for this, we imputed missing data using a sophisticated on the use of a computing device to estimate the probability of
approach based on a global dataset of 12 international myocardial infarction (International Publication Number
cohorts. Seventh, data on quality of symptoms and WO2022043229A1). AZ is scientific director and CEO of Cardio-CARE,
symptom characteristics and the exact type of detected SB is scientific advisor of Cardio-CARE, HS was a biostatistician at
Cardio-CARE, and EDC is a biostatistician at Cardio-CARE, a 100% non-
ischaemic ECG signs were not available. Digital or sec- profit daughter company of the Kühne Foundation. Cardio-CARE is a
ondarily digitised ECGs were also not available in both shareholder of ART-EMIS Hamburg. SB receives consulting fees from
cohorts. The integration of digital ECGs rather than Thermo Fisher, and payment or honoraria from Abbott Diagnostics,
dichotomised ECG data might be achieved in the future. Amarin, AMGEN, AstraZeneca, Bayer, Bristol Meyers Squibb,
Boehringer Ingelheim, Daiichi Sankyo, GSK, LumiraDx, Novartis, Roche
In conclusion, the patient-tailored ARTEMIS algorithm Diagnostics, and Thermo Fisher. SB is a member of the advisory board
applied with a single POC hs-cTnI test allows for rapid, of Thermo Fisher. SB and AZ were co-founders of ARTEMIS in 2022.
safe, but substantially more efficient direct rule-out of Both sold their shares in 2023 and have no personal financial or legal
myocardial infarction than guideline-recommended relationship to ARTEMIS. KS declares no competing interests.
pathways—even across distinct clinical subgroups. If Data sharing
implemented in clinical practice as a medical decision All individual-level data of the included cohorts can be shared upon
reasonable request to the corresponding author and completion of data
support tool, the ARTEMIS-POC algorithm holds transfer agreement forms. The algorithm is proprietary and subject to
promise for expediting early discharge from the patent application. However, we can share it for research purposes based
emergency department, and might extend safe and rapid on a written agreement and a request made to the corresponding author.
triage of patients with suspected myocardial infarction to Acknowledgments
diverse settings, including pre-hospital, geographically The authors thank the technical staff involved in SEIGE and SAMIE
responsible for sample processing and laboratory analyses. The authors
isolated, and ambulatory care settings.
also thank the patients who participated in this study for their
Collaborators commitment.
SEIGE and SAMIE investigators: Laura Stephensen, Emily Brownlee,
References
Ellyse McCormick, Gavin Fincher, Emma J Hall, Rebecca Hancock,
1 Blomkalns AL, Gibler WB. Chest pain unit concept: rationale and
Niranjan Gaikwad, Vinay Gangathimmaiah, Christian Hamilton-Craig,
diagnostic strategies. Cardiol Clin 2005; 23: 411–21.
Andrew Hobbins-King, Gerben Keijzers, Maryam Khorramshahi Bayat,

www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2 9

 Articles

2 Neumann JT, Sörensen NA, Westermann D. Biomarkers in the 15 Thygesen K, Alpert JS, Jaffe AS, et al. Fourth universal definition of
triage of chest pain: are we making progress? Biomarkers Med 2016; myocardial infarction (2018). Eur Heart J 2018; 39: 3757–58.
10: 345–47. 16 Hartikainen TS, Sörensen NA, Goßling A, et al. Application of the
3 Greenslade J, Cho E, Van Hise C, et al. Evaluating rapid rule-out of Fourth Universal Definition of MI using FDA-recommended sex-
acute myocardial ınfarction using a high-sensitivity cardiac troponin specific troponin cutoff concentrations. J Am Coll Cardiol 2021;
I assay at presentation. Clin Chem 2018; 64: 820–29. 77: 2346–48.
4 Sandoval Y, Lewis BR, Mehta RA, et al. Rapid exclusion of acute 17 Bhuiya FA, Pitts SR, McCaig LF. Emergency department visits for
myocardial ınjury and ınfarction with a single high-sensitivity chest pain and abdominal pain: United States, 1999–2008.
cardiac troponin T in the emergency department: a multicenter NCHS Data Brief 2010: 1–8.
united states evaluation. Circulation 2022; 145: 1708–19. 18 Pedersen CK, Stengaard C, Bøtker MT, Søndergaard HM, Dodt KK,
5 Pickering JW, Than MP, Cullen L, et al. Rapid rule-out of acute Terkelsen CJ. Accelerated rule-out of acute myocardial infarction
myocardial ınfarction with a single high-sensitivity cardiac troponin using prehospital copeptin and in-hospital troponin: the AROMI
T measurement below the limit of detection: a collaborative meta- study. Eur Heart J 2023; 44: 3875–88.
analysis. Ann Intern Med 2017; 166: 715–24. 19 Januzzi JL Jr, McCarthy CP. Evaluating chest pain in the emergency
6 Byrne RA, Rossello X, Coughlan JJ, et al. 2023 ESC guidelines for department: searching for the optimal gatekeeper. J Am Coll Cardiol
the management of acute coronary syndromes. Eur Heart J 2023; 2018; 71: 617–19.
44: 3720–826. 20 Abe T, Olanipekun T, Adedinsewo D, et al. Trends and outcomes of
7 Kontos MC, de Lemos JA, Deitelzweig SB, et al. 2022 ACC expert ST-segment-elevation myocardial ınfarction among young women
consensus decision pathway on the evaluation and disposition of in the United States. J Am Heart Assoc 2023; 12: e026811.
acute chest pain in the emergency department: a report of the 21 Than MP, Pickering JW, Sandoval Y, et al. Machine learning to
American College of Cardiology solution set oversight committee. predict the likelihood of acute myocardial ınfarction. Circulation
J Am Coll Cardiol 2022; 80: 1925–60. 2019; 140: 899–909.
8 Lowry MTH, Doudesis D, Boeddinghaus J, et al. Troponin in early 22 Doudesis D, Lee KK, Boeddinghaus J, et al. Machine learning for
presenters to rule out myocardial infarction. Eur Heart J 2023; diagnosis of myocardial infarction using cardiac troponin
44: 2846–58. concentrations. Nat Med 2023; 29: 1201–10.
9 Lowry MTH, Doudesis D, Wereski R, et al. Influence of age on the 23 Boeddinghaus J, Doudesis D, Lopez-Ayala P, et al. Machine
diagnosis of myocardial ınfarction. Circulation 2022; 146: 1135–48. learning for myocardial ınfarction compared with guideline-
10 Twerenbold R, Badertscher P, Boeddinghaus J, et al. 0/1-hour triage recommended diagnostic pathways. Circulation 2024;
algorithm for myocardial infarction in patients with renal 149: 1090–101.
dysfunction. Circulation 2018; 137: 436–51. 24 Herman R, Meyers HP, Smith SW, et al. International evaluation of
11 Wahrenberg A, Magnusson PK, Discacciati A, et al. Family history an artificial intelligence-powered electrocardiogram model detecting
of coronary artery disease is associated with acute coronary acute coronary occlusion myocardial infarction.
syndrome in 28,188 chest pain patients. Eur Heart J Digit Health 2023; 5: 123–33.
Eur Heart J Acute Cardiovasc Care 2019; 9: 741–47. 25 Collinson P, Aakre KM, Saenger A, et al. Cardiac troponin
12 Neumann JT, Twerenbold R, Ojeda F, et al. Personalized diagnosis in measurement at the point of care: educational recommendations on
suspected myocardial infarction. Clin Res Cardiol 2023; 112: 1288–301. analytical and clinical aspects by the IFCC Committee on Clinical
13 Cullen L, Collinson PO, Giannitsis E. Point-of-care testing with Applications of Cardiac Bio-Markers (IFCC C-CB). Clin Chem Lab Med
high-sensitivity cardiac troponin assays: the challenges and 2023; 61: 989–98.
opportunities. Emerg Med J 2022; 39: 861–66.
14 Apple FS, Smith SW, Greenslade JH, et al. Single high-sensitivity
point-of-care whole-blood cardiac troponin I measurement to rule out
acute myocardial ınfarction at low risk. Circulation 2022; 146: 1918–29.

10 www.thelancet.com/digital-health Published online August 29, 2024 https://doi.org/10.1016/S2589-7500(24)00191-2