Machine-Learning Troponin Scores: What CoDE-ACS Can Rule Out

Chest pain is one of the most common reasons adults come to the emergency department, and most of those patients do not have a myocardial infarction. High-sensitivity troponin pathways already send many of them home after one or two draws. A newer class of tools runs the troponin value, the patient's age and sex, and a handful of clinical variables through a machine-learning model and returns a single probability from 0 to 100. The validation numbers are striking, and the tools are inching toward the electronic record.

Bottom line: treat a machine-learning MI probability as a second rule-out read that is only valid when your lab runs the same troponin assay the model was trained on, the patient has no ST elevation, and your own data confirm the negative predictive value before any score drives disposition. Until a prospective impact trial reports, the ECG, the history, and your pathway still decide.

What the Models Actually Compute

CoDE-ACS (Collaboration for the Diagnosis and Evaluation of Acute Coronary Syndrome) uses extreme gradient boosting, chosen after it outperformed logistic regression, naive Bayes, and random forest [1]. It takes a high-sensitivity cardiac troponin I concentration as a continuous value plus 12 variables: age, sex, hours from symptom onset, chest pain, known ischemic heart disease, hyperlipidemia, heart rate, systolic blood pressure, Killip class, ischemia on the ECG, eGFR, and hemoglobin [1]. Separate models handle patients with and without myocardial injury at presentation, defined as troponin above the sex-specific 99th percentile of 16 ng/L in women and 34 ng/L in men [1].

The outcome was an adjudicated type 1, 4b, or 4c MI during the index admission under the Fourth Universal Definition, with STEMI excluded [1]. The developers fixed a score under 3 as low probability, chosen to hit a negative predictive value of 99.5% and a sensitivity of 90% in patients without injury, and a score of 61 or more as high probability [1].

MI3 (myocardial-ischemic-injury-index), from a separate collaboration, is leaner: age, sex, and paired high-sensitivity troponin I values, also run through gradient boosting [3].

How Many More Patients Go Home

CoDE-ACS was trained on 10,038 patients (48% women) from the High-STEACS trial in 10 Scottish hospitals and externally validated on 10,286 patients (35% women) from seven prospective cohorts in Switzerland, Spain, Poland, the Czech Republic, Australia, New Zealand, and the United States [1]. Discrimination was high, with an AUC of 0.953 (95% CI 0.947 to 0.958) on a single draw and 0.966 with serial testing [1].

At presentation, 61% of the validation cohort scored low probability, with a negative predictive value of 99.6% and a sensitivity of 97.9% [1]. A fixed rule-out of troponin under 5 ng/L, applied only to patients without ECG ischemia who presented more than 3 hours after onset, ruled out 27% with a negative predictive value of 99.7% and a sensitivity of 98.3% [1]. On the rule-in side, a score of 61 or more flagged 10% of patients with a positive predictive value of 75.5%, against 16% and 63.6% for the 99th percentile [1].

The low-probability group also did well after discharge: cardiac death was 0.1% at 30 days and 0.3% at 1 year, compared with 0.5% and 2.8% in the intermediate group and 1.8% and 4.2% in the high group [1].

Against the 0/1-Hour Pathway

Boeddinghaus and colleagues tested CoDE-ACS in the APACE cohort: 4,105 patients without ST elevation at 12 sites in 5 countries, with troponin drawn at 0, 1, and 2 hours; 575 (14%) had a type 1 MI [2]. At presentation, CoDE-ACS called 56% low probability with a negative predictive value of 99.7% and a sensitivity of 99.0%, versus 25% for the ESC 0-hour rule-out and 35% for the High-STEACS pathway [2].

After a second draw, CoDE-ACS classified 65% (1 hour) or 68% (2 hours) as low probability, with a negative predictive value of 99.7% [2]. The ESC 0/1-hour and 0/2-hour pathways ruled out 49% and 53% with a negative predictive value of 100%. High-STEACS ruled out 71% at 99.7% [2]. The guideline pathways left 31% and 28% of patients in the observation zone, against 16% and 14% for CoDE-ACS [2].

The tradeoff is visible in those numbers. The model rules out more patients at presentation and leaves fewer in limbo, but the ESC pathways still post a slightly higher negative predictive value.

MI3 and the Threshold Problem

MI3 was trained on 3,013 patients and tested on 7,998, with MI rates of 13.4% and 10.6% [3]. AUC in the test set was 0.963, similar in early and late presenters [3]. A low-risk threshold of 1.6, chosen in the training set for 99% sensitivity, captured 69.5% of test patients with a negative predictive value of 99.7%, but sensitivity fell to 97.8% [3]. A high-risk threshold of 49.7 flagged 10.6% with a positive predictive value of 71.8% [3].

That drop from a 99% target to 97.8% in new patients is the general lesson. A threshold tuned in one population drifts when it meets another, and the cutoff is a local decision about how many missed MIs a department will accept.

Where the Evidence Stops

The subgroup data are uneven. In CoDE-ACS validation, low-probability performance was similar in 3,629 women and 6,657 men, but the high-probability positive predictive value was 67.5% in women versus 78.5% in men [1]. False negatives stayed under 1 in 100 across subgroups, and the authors still call for prospective validation in patients over 65 and those with ischemic heart disease or renal impairment [1]. Most participants were White, so performance across ethnic groups could not be assessed [1].

Two limits matter most for a US department. CoDE-ACS was built on a single manufacturer's assay, the Abbott ARCHITECT high-sensitivity troponin I, and the authors state that other assays require retraining and validation [1]. In addition, every validation cohort had already finished enrolling, and no clinician's care was guided by the score; the authors and the APACE team both say a prospective impact study is still needed [1,2].

Key Takeaways

• CoDE-ACS called 61% of 10,286 validation patients low probability on one draw, with a negative predictive value of 99.6%, versus 27% for a fixed troponin rule-out [1].
• After serial draws, the ESC 0/1-hour pathway still had a higher negative predictive value (100% versus 99.7%) but left 31% in observation, against 16% for CoDE-ACS [2].
• CoDE-ACS was trained and validated only on the Abbott ARCHITECT hs-cTnI assay; other assays need retraining [1].
• MI3's low-risk threshold, set for 99% sensitivity in training, delivered 97.8% in 7,998 new patients [3].

FAQ

Can a machine-learning troponin score rule out myocardial infarction on its own?

Not yet in practice. CoDE-ACS reached a negative predictive value of 99.6% on a single draw in 10,286 patients, but no validation cohort had care guided by the score, and the authors call for prospective impact testing [1]. It excludes STEMI and needs the ECG and clinical variables as inputs.

Does CoDE-ACS work with any high-sensitivity troponin assay?

No. It was trained and validated on the Abbott ARCHITECT high-sensitivity troponin I assay, and the developers state that other assays require retraining and validation because troponin assays are not standardized [1].

How does CoDE-ACS compare with the ESC 0/1-hour algorithm?

In 4,105 APACE patients, CoDE-ACS ruled out 65% after a 1-hour draw with a negative predictive value of 99.7%, while the ESC 0/1-hour pathway ruled out 49% at 100% [2]. CoDE-ACS left 16% in the observation zone versus 31% [2].

Dr. Chet's Take

I have spent more than 25 years as an emergency physician, and chest pain is a complaint where the evidence has moved faster than almost anywhere else in our specialty. High-sensitivity troponin pathways changed the conversation from "rule out with a long stay" to "rule out in an hour or two." This article is right that a probability score is a real step forward in how that information is packaged. A single number that blends troponin with age, timing, and the ECG is closer to how experienced physicians actually think. What I would add is that a probability is only useful if someone in the department has decided, in advance and in writing, what number sends a patient home and what number keeps them. Without that decision, a score becomes one more data point that nobody owns.

That being said, the assay limitation is the one most likely to be ignored. Labs change troponin platforms for cost and contracting reasons, often without the emergency department at the table. A model that quietly runs on the wrong assay will look exactly like one that runs on the right assay until a missed MI shows up in peer review. I also want to see the prospective impact trial before any department lets a score drive discharge by itself. Retrospective accuracy is the beginning of the evidence. How clinicians behave when the number is on the screen is a separate question, and it has not been answered for this tool. Anchoring on a reassuring number is a human failure mode, and the model cannot correct for it.

If you are the emergency physician seeing the patient, use the score only alongside your ECG, your history, and your existing pathway, and document why you agreed or disagreed with it. If you lead the department, confirm your lab's assay against the model's training assay, then run the score silently on your own patients for several months before anyone sees it. If you chair the chest pain committee, set the low-probability cutoff from your local miss rate and put the review date in the policy. Ask your lab director to tell the department before any troponin platform change, and treat that change as a reason to suspend the score until it is revalidated. A probability that nobody has checked against local outcomes is an unvalidated test.

— Chester Shermer, MD, FACEP | Emergency Medicine, 25+ Years Clinical Experience | State Surgeon

Continue Your Training

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download the AI in EM Survival Guide here.

Browse All Courses at Global MedOps Command

Relevant Reading on Global MedOps Command:

How to Avoid Becoming an AI Casualty — Dr. Shermer's guide to navigating AI tools in clinical and operational settings without compromising judgment or patient outcomes.

Emergency Department Efficiency Playbook — Practical systems for throughput, triage optimization, and operational efficiency.

Read more from Dr. Shermer on Medium

Connect with Dr. Shermer: LinkedIn — Chester "Chet" Shermer, MD, FACEP.

References

1. Doudesis D, Lee KK, Boeddinghaus J, et al. Machine learning for diagnosis of myocardial infarction using cardiac troponin concentrations. Nat Med. 2023;29(5):1201-1210. PubMed 37169863
2. Boeddinghaus J, Doudesis D, Lopez-Ayala P, et al. Machine learning for myocardial infarction compared with guideline-recommended diagnostic pathways. Circulation. 2024;149(14):1090-1101. PubMed 38344871
3. Than MP, Pickering JW, Sandoval Y, et al. Machine learning to predict the likelihood of acute myocardial infarction. Circulation. 2019;140(11):899-909. PubMed 31416346

Related simulation training: EM-Sim · Books by Dr. Shermer