Details

Autor(en) / Beteiligte

• Lopez Ayala, P
• Boeddinghaus, J
• Nestelberger, T
• Strebel, I
• Koechlin, L
• Rubini Gimenez, M
• Miro, O
• Martin-Sanchez, F J
• Keller, D
• Mueller, C

Titel

External validation of the myocardial-ischemic-injury-index: a machine learning algorithm for the early diagnosis of myocardial infarction

Ist Teil von

• European heart journal, 2023-11, Vol.44 (Supplement_2)

Erscheinungsjahr

2023

Quelle

Oxford Journals 2020 Medicine

Beschreibungen/Notizen

• Abstract Introduction The myocardial-ischemic-injury-index (MI3) is a machine learning algorithm derived in 2019 for the early diagnosis of type 1 non-ST-elevation myocardial infarction (NSTEMI). MI3 incorporates age, sex, and two high-sensitivity cardiac troponin I (hs-cTnI) Architecht measurements. The performance of the MI3 algorithm is unknown in a setting where the time difference between the first and second troponin measurement is less than 3 hours. Thus, external validation seems mandatory before this algorithm could be considered for clinical use. Purpose To externally validate the performance of the MI3 algorithm and compare its performance with the ESC 0/1h-algorithm, the recommended algorithm by the European Society of Cardiology NSTEMI guidelines. Methods In a multicentre international diagnostic study, we prospectively enrolled unselected patients presenting to the emergency department with symptoms suggestive of MI. Final diagnoses were centrally adjudicated by two independent cardiologists using all available medical records, including cardiac imaging, serial hs-cTn and 90-days follow-up. Hs-cTnI (Architect) concentrations were measured in a blinded fashion at presentation and serially thereafter. The primary endpoint was type 1 NSTEMI during the index visit. Results Among 6487 patients with two available hs-cTnI concentrations, type 1 NSTEMI was the adjudicated final diagnosis in 882 (13.6%) patients. The median time difference between blood draws was 60 minutes (IQR: 57-70). Model performance was good, with a very high area under the receiver-operating-characteristic curve (0.961 [95% CI: 0.957-0.965]) and an optimal calibration overall (intercept -0.09 [-0.2-0.02]; slope 1.02 [0.97-1.08]), Figure 1. The originally defined low-probability MI3 score<1.6 and high-probability MI3 score≥49.7 triaged 4186 (64.5%) patients towards rule-out (sensitivity 99.1% [98.2-99.5]; negative predictive value [NPV] 99.8% [99.6-99.9]) and 915 (14.1%) patients towards rule-in (specificity 95.0% [94.3-95.5]; positive predictive value [PPV] 69.1% [66.0-72.0]), respectively (Figure 2). Both the sensitivity and the NPV of the ESC 0/1h-algorithm were greater than of the MI3 algorithm (difference for sensitivity: 0.84% [0.18-1.51], P=0.008; difference for NPV: 0.18% [0.05-0.33], P=0.016). In contrast, the specificity and PPV for the MI3 algorithm were higher (difference for specificity: 3.85% [3.29-4.40], P<0.001; difference for PPV: 7.82% [5.9-9.74], P<0.001). Conclusion In patients with two hs-cTnI (Architect) measurements with a median time difference of 60 minutes, the MI3 algorithm showed a very high discrimination and good calibration for type 1 NSTEMI.Diagnostic performance MI3 algorithmCutoff performance of the MI3 algorithm