External validation of the myocardial-ischemic-injury-index: a machine learning algorithm for the early diagnosis of myocardial infarction
Key result
MI3 yields ~0.96 AUC for type 1 NSTEMI, trading sensitivity for higher specificity vs ESC 0/1h.
- AUC 0.961
- 95% CI 0.957-0.965
- n=6,487
Why the study?
The performance of the MI3 machine learning algorithm was unknown when the interval between serial troponin measurements was less than 3 hours, requiring external validation before clinical use.
Does the MI3 algorithm accurately diagnose type 1 NSTEMI compared to the ESC 0/1h-algorithm in patients presenting with symptoms suggestive of MI?
Population
6487 unselected ED patients presenting with symptoms suggestive of MI with two available hs-cTnI concentrations
Comparison
MI3 machine learning algorithm vs ESC 0/1h-algorithm
Design
Multicentre international prospective diagnostic study
Follow-up
90 days
Authors
Discussion
Loading...
Member takes
Key expert perspectives
Captured external expert commentary on this paper, strongest first. Original sources are linked where available.
SGSteve GoodacreProfessor of Emergency Medicine, University of Sheffield
“This intriguing study shows how AI can use complex analysis, rather than a simple rule, to improve diagnosis. This doesn't (yet) show that we can replace doctors with computers. Experienced clinicians know that diagnosis is a complex business. Indeed, the 'ground truth' used to judge whether the AI algorithm was accurate was a judgement made by clinicians.”
Implication
MI3 shows strong performance at 1-hour hs-cTnI intervals; leaves open accelerated pathways pending prospective outcome trials.
Key Points
- To externally validate the machine learning-based myocardial-ischemic-injury-index (MI3) algorithm for early diagnosis of type 1 NSTEMI and compare its diagnostic performance to the ESC 0/1-hour algorithm.
- Prospective, multicenter international diagnostic study enrolling unselected emergency department patients presenting with symptoms suggestive of myocardial infarction (N=6,487 with two serial blood draws).
- High-sensitivity cardiac troponin I (hs-cTnI, Architect) was measured blindly at presentation and serially with a median interval of 60 minutes (IQR: 57–70).
- Type 1 NSTEMI diagnosis was centrally adjudicated by two independent cardiologists using serial hs-cTn, cardiac imaging, and 90-day follow-up.
- Among 6,487 patients (882 [13.6%] adjudicated type 1 NSTEMI), the MI3 algorithm demonstrated an AUC of 0.961 (95% CI: 0.957–0.965), calibration intercept of -0.09 (95% CI: -0.2–0.02), and slope of 1.02 (95% CI: 0.97–1.08).
- Low-probability MI3 score (<1.6) ruled out 4,186 (64.5%) patients with sensitivity of 99.1% (95% CI: 98.2–99.5) and NPV of 99.8% (95% CI: 99.6–99.9); high-probability score (≥49.7) ruled in 915 (14.1%) patients with specificity of 95.0% (95% CI: 94.3–95.5) and PPV of 69.1% (95% CI: 66.0–72.0).
- Compared with MI3, the ESC 0/1h-algorithm had higher sensitivity (difference: 0.84% [95% CI: 0.18–1.51], P=0.008) and NPV (difference: 0.18% [95% CI: 0.05–0.33], P=0.016), but MI3 achieved higher specificity (difference: 3.85% [95% CI: 3.29–4.40], P<0.001) and PPV (difference: 7.82% [95% CI: 5.9–9.74], P<0.001).
Study Design
Type
Observational (n=6,487)
Blinding
Blinded
Multicenter
Yes
Structured PICO
Does the MI3 algorithm accurately diagnose type 1 NSTEMI compared to the ESC 0/1h-algorithm in patients presenting with symptoms suggestive of MI?
P
Population
6,487 unselected patients presenting to the emergency department with symptoms suggestive of myocardial infarction, followed for 90 days.
E
Exposure
Myocardial-ischemic-injury-index (MI3) machine learning algorithm (incorporating age, sex, and two hs-cTnI Architect measurements).
C
Comparator
ESC 0/1h-algorithm.
O
Outcome
Type 1 NSTEMI during the index visit.hard clinical
Main Result
Effect estimate: AUC 0.961 (95% CI 0.957-0.965)
The MI3 machine learning algorithm provides excellent discrimination for early diagnosis of type 1 NSTEMI, with higher specificity and positive predictive value but slightly lower sensitivity than the standard ESC 0/1h-algorithm.
Cite This Study
Boeddinghaus et al. (2023) conducted an observational in Myocardial Infarction (n=6,487). Myocardial-ischemic-injury-index (MI3) algorithm vs. ESC 0/1h-algorithm was evaluated on Type 1 NSTEMI during the index visit (AUC 0.961, 95% CI 0.957-0.965). The MI3 algorithm demonstrated very high discrimination for type 1 NSTEMI (AUC 0.961; 95% CI 0.957-0.965) and had higher specificity but lower sensitivity compared to the ESC 0/1h-algorithm.