ECG-SMART model detected occlusion myocardial infarction with moderate accuracy (AUC 0.74), outperforming standard ECG features, especially within 3h of chest pain onset (AUC 0.78).
Does the ECG-SMART model improve the diagnosis of occlusion myocardial infarction compared to standard ECG features in patients presenting to the emergency department with symptoms suggestive of MI?
In ED patients with symptoms suggestive of MI, the ECG-SMART machine learning model showed moderate discrimination for detecting occlusion myocardial infarction, outperforming standard ECG features.
Absolute Event Rate: 0% vs 0%
Abstract Introduction The ECG-SMART model is an electrocardiogram (ECG)-based machine learning algorithm derived in 2023 for the early diagnosis of occlusion myocardial infarction (OMI). It was derived using 12-lead pre-hospital ECGs and showed very high diagnostic performance in derivation and validation cohorts from the United States. However, its performance in the emergency department (ED) setting, where the time difference between chest pain onset and ECG recording is longer, remains unknown. Therefore, external validation seems mandatory before considering its clinical implementation in the ED. Purpose To externally validate the performance of the ECG-SMART model and compare its performance with standard ischemic ECG features. 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. The primary endpoint was OMI, defined as the angiographic evidence of an acute culprit lesion with a thrombolysis in myocrdial infarction (TIMI) flow grade of 0-1 or a TIMI flow of 2 and a peak hs-cTnT ≥500 ng/L. TIMI flow grades were assessed from coronary angiography reports and images by two independent interventional cardiologists. Results Among 3797 patients with an available digital 12-lead ECG, OMI was the adjudicated final diagnosis in 249 (6.6%) patients. The median time difference between chest pain onset and ED presentation was 6.5 hours (IQR: 2.5-24.0). Patients with OMI had a higher probability score (24.5 vs 10.6, Figure 1a). Model discrimination was moderate, with an overall area under the receiver-operating-characteristic curve of 0.74 (95% CI: 0.71-0.78), being superior to ST-segment elevation (AUC 0.65), ST-segment depression (AUC 0.66) or T wave inversion (AUC 0.56), Figure 1b. Discrimination was larger in early presenters (chest pain onset ≤3h; AUC 0.78 0.72-0.84) and lower in late presenters (defined as chest pain onset 12h; AUC 0.69 0.60-0.78). No differences were observed according to sex or age. The calibration plot showed an overestimation of OMI risk (intercept -0.98 -1.12-0.85; slope 0.90 0.77-1.04). After recalibration of the intercept predicted risks were comparable to observed risks (Figure 2). The originally defined high-probability score ≥68 triaged 24 (0.6%) patients towards rule-in (specificity 99.8% 99.6-99.9; positive predictive value PPV 75.0% 55.1-88.0). A lower probability score (≥50) resulted in similar performance to the derivation cohort (2.3% rule-in, specificity 98.7% and PPV 48.3%). Conclusion In ED chest pain patients with a median time since chest pain onset of 6.5 hours, ECG-SMART showed moderate discrimination for detecting OMI. The performance of the machine learning model improved among patients with shorter chest pain onset and was superior to standard ECG features.Fig1.Diagnostic Performance Fig2.Calibration plots
Ayala et al. (Sat,) reported a other. ECG-SMART model detected occlusion myocardial infarction with moderate accuracy (AUC 0.74), outperforming standard ECG features, especially within 3h of chest pain onset (AUC 0.78).