Comparison of ordinal AI-ECG versus ordinal hs-cTnT logistic regression for diagnosis of occlusion myocardial infarction: a prospective multicenter validation study
I Strebel, P Lopez Ayala, T Zimmermann, W Zeng, S Knecht, E Kaplan, A Champetier, K Arslani, M Wagner, M Christ, O Miro, G Leibundgut, F Mahfoud, C Mueller, J BoeddinghausAbstract
Background
In the occlusion myocardial infarction (OMI) paradigm, the focus shifts from rigid electrocardiogram (ECG) criteria to the underlying pathology of acute coronary occlusion requiring urgent reperfusion. Direct comparison of an ordinal AI-ECG model against an ordinal troponin-based logistic regression model for OMI/NOMI detection in unselected emergency department (ED) patients at presentation is lacking.
Methods
This study fine-tuned an ordinal AI-ECG model from a foundation model (pre-trained on pooled cohorts n=10.7 million), with preliminary fine-tuning on retrospective cohorts (n=338’580) using ICD-10 codes and final fine-tuning on pooled cohorts (n=13’879) with ordinal OMI/NOMI outcomes. An ordinal logistic regression model using hs-cTnT (LR-Trop) as predictor was fitted in parallel. Models were externally validated and compared in a prospective multicenter study of unselected ED patients with chest pain (validation n=4728), with ECG and hs-cTnT at presentation. Diagnoses were centrally adjudicated. Primary outcome was Occlusion Myocardial infarction (angiographic culprit with TIMI 0–1, or TIMI 2–3 with peak hs-cTnT ≥1000 ng/L). TIMI flow was assessed by two cardiologists. Performance was evaluated by receiver operating characteristic curves (ROC) and precision-recall curves (PR). A sensitivity analysis was performed in the combined group of patients with NSTEMI-OMI and those without MI.
Results
Among 4728 patients, occlusion myocardial infarction (OMI) was the final diagnosis in 300 (6.3%), of whom 122 (41%) had STEMI. Model discrimination for OMI was high (AI-ECG AUC 0.89 [0.87-0.90] vs LR-Trop 0.93 [0.92-0.95]). For a threshold with PPV of 0.7, the proportion of patients with OMI identified was similar (AI-ECG 0.35 [0.30-0.41] vs LR-Trop 0.41 [0.35-0.47], p >0.05) (Figures A and C). In patients with NSTEMI-OMI or no MI, model discrimination dropped for AI-ECG while it remained high for LR-Trop (AI-ECG AUC 0.84 [0.81-0.87] vs LR-Trop 0.93 [0.91-0.95]). For a threshold with PPV of 0.7, the proportion of patients with NSTEMI-OMI identified was significantly lower for AI-ECG compared to LR-Trop (0.07 [0.04-0.11] vs 0.26 [0.20-0.33], p < 0.05) (Figures B and D).
Conclusions & clinical implications
In this prospective multicenter validation of unselected emergency department patients with chest pain, the fine-tuned ordinal AI-ECG model demonstrated strong discrimination for OMI and identified a comparable proportion of cases to the hs-cTnT-based LR-Trop. While the performance of AI-ECG was high for OMI, it was inferior to ordinal logistic regression with cardiac troponin testing in patients with NSTEMI-OMI, highlighting the need for optimised models in this vulnerable population.Figures A-DGraphical Abstract