External validation of an electrocardiogram-based convolutional neural network model for the early diagnosis of occlusion myocardial infarction

Background The Queen of Hearts (QoH) model is an electrocardiogram (ECG)-based convolutional neural network (CNN) model for the early diagnosis of occlusion myocardial infarction (OMI), showing good diagnostic performance. Because model derivation was performed in a retrospective cohort of patients presenting to a cardiovascular facility with acute coronary syndrome, external validation in an emergency department (ED) setting in unselected patients is lacking. Purpose To externally validate the performance of the QoH model and compare it with standard ischemic ECG features and high-sensitivity cardiac troponin T (hs-cTnT). Methods In a multicentre international diagnostic study, we prospectively enrolled unselected patients presenting to the ED with symptoms suggestive of myocardial infarction. For this analysis, patients were excluded if a digital ECGs was not available. Final diagnoses were centrally adjudicated by two independent cardiologists using all available medical records. The primary endpoint was OMI, defined as angiographic evidence of an acute culprit lesion with a TIMI flow 0-1, or TIMI flow 2-3 with a peak hs-cTnT ≥1000 ng/L. TIMI flow grades were assessed from coronary angiography reports and images by two interventional cardiologists. Model performance was assessed by means of discrimination (area under the receiver operating characteristic curve) and calibration. Results Among 4728 patients with an available digital ECG, OMI was the final diagnosis in 300 (6.3%) patients. Patients with OMI had a higher probability score (40.4% vs 0.12%, Figure 1a). Model discrimination was high (AUC 0.88; 95% CI:0.85-0.90), outperforming ischemic ECG features, Figure 1b. The calibration plot showed risk underestimation at low and overestimation at high predicted probabilities, improving after recalibration, Figure 2. When QoH was added to a rich-for-prior-information model including hs-cTnT, discrimination was improved (AUC base model: 0.94 [0.93-0.95] vs extended model: 0.96 [0.95-0.97]). The originally derived rule-in probability ≥50% triaged 222 (4.7%) patients towards rule-in (specificity 98.2% [97.8-98.5]; positive predictive value [PPV] 64.0% [57.5-70.0]), resulting in a better performance compared to the European Society of Cardiology rule-in hs-cTnT cutoff ≥52 ng/L (specificity 93.1% [92.4-93.8]; PPV 41.8% [37.7-46.1]). Conclusions In ED patients with chest pain, an ECG-based CNN model demonstrated high diagnostic accuracy for detecting OMI, outperforming traditional ECG features of ischemia and providing added value to hs-cTnT. Rule-in performance was good showing superior accuracy for detecting OMI compared to the established ESC single-measurement hs-cTnT strategy.