Abstract 4145357: Detecting ST Elevation Myocardial Infarction on a Noisy Single Limb-lead ECG: An Artificial Intelligence-enabled Approach Adaptable to Portable Devices

Background: While wearable and portable devices recording 1-lead ECGs can potentially enable accessible cardiovascular disease detection, most capture a single limb lead (lead I) that is clinically insufficient to define STEMI. Moreover, the acquisition is noisy. We sought to evaluate if an artificial intelligence (AI) algorithm can detect STEMI and its subtypes based on location (anterior, inferior, lateral, posterior) from noisy 1-lead ECG tracings. Methods: We extracted lead I data from ECGs in patients at a large and diverse US health system (2015-2023) to train a multi-label convolutional neural network (CNN) for detecting STEMI and its different subtypes. To ensure that the CNN identified the STEMI signature independent of associated findings, every ECG with STEMI (case) in the training set was matched to 10 ECGs without STEMI (controls) but with the same age profile, sex, and other rhythm and conduction disorders. Each ECG recording was augmented with random Gaussian noise during training to mimic real-world noises of specific frequencies, including muscle tremors, electrode motion, and others. The model was evaluated in a held-out test set of lead I tracings from ECGs recorded during ED visits for patients with chest pain. Results: We used 49,995 lead I ECGs from 39,697 individuals (65 years [IQR, 52-76], 18,926 [48%] women, 5,941 [15%] Black) for model development. The held-out test set included one ECG each from 7551 patients, including 104 (1.4%) with STEMI (84 anterior, 103 inferior, 72 lateral, 7 posterior STEMI). The model achieved high discrimination (AUROC of 0.828 [95% CI, 0.777-0.875]) for detecting STEMI from lead I ECGs in the test set, with AUROCs of 0.734 (0.656-0.802) for anterior, 0.769 (0.726-0.818) for inferior, 0.849 (0.787-0.889) for lateral, and 0.747 (0.518-0.968) for posterior STEMI. Overall, the model had a sensitivity of 91.5% (85.6-96.9), and a specificity of 43.0% (41.8-43.9), with consistent performance among demographic subgroups. Conclusion: A noise-adapted AI model using only 1-lead ECG tracings was able to identify various subtypes of STEMI with moderate to high accuracy. Using portable ECG devices and smartwatches, an AI-ECG approach can enable first-line triaging in individuals with chest pain, facilitating early detection of STEMI in the community and high-acuity settings.