[2410.04133] An Electrocardiogram Foundation Model Built on over 10 Million Recordings with External Evaluation across Multiple Domains

View a PDF file from the paper entitled Electrical Planning Form based on more than 10 million records with an external evaluation through multiple fields, by Jun Li and 8 other authors

PDF HTML (experimental) view

a summary:artificial intelligence (AI) has shown great potential for ECG analysis and evaluation of cardiovascular diseases. Recently, basic models have played a great role in medical artificial intelligence. The development of the ECG Foundation model carries a promise to raise AI-ECG research to new horizons. However, the construction of such a model faces many challenges, including insufficient database sample sample and insufficient generalization across multiple areas. In addition, there is a noticeable performance gap between monoch and multi -lead -for -scoring ECG analyzes. We presented an ECGFONSER foundation, a model for general purposes that enhances ECG clarifications in the real world from heart disease experts to expand the diagnostic capabilities of ECG analysis. ECGFONSER has been trained on more than 10 million ECGs with 150 stickers from the Harvard-EMory ECG database, allowing diagnosis of comprehensive cardiovascular disease through ECG analysis. The model is designed to be an effective solution outside the box, and it is well set for the estuary tasks, which increases the ability to use. More importantly, we have expanded its application to the least -ranking heart planning, and ECGS in particular. ECGFOUNER applies to support various estuaries in mobile surveillance scenarios. Experimental results show that ECGFONSER achieves experts level on internal verification groups, as Auroc exceeds 0.95 to diagnose eighty. It also shows the performance of a strong classification and generalization through various diagnoses on external health verification groups. When adjusted, Ecfonser surpasses the basic models of demographic analysis, detecting clinical events, and diagnosing the cross heart rhythm. The form and data trained publicly will be released when posting through this URL http. Our symbol is available in this URL https