AI Model Detects Heart Failure From ECG With AUC Up to 0.96 in Norway Study

Norwegian researchers trained an open-source AI model on 284,000 ECGs to detect heart failure across the full ejection fraction spectrum, including the commonly missed HFpEF type. The team used a method called pragmatic labelling, combining ICD diagnostic codes with NT-proBNP biomarker data to improve training accuracy. In prospective testing on over 43,000 patients, the model achieved an overall AUC of 0.84, rising to 0.91 for HFrEF and up to 0.96 under strict labelling conditions. Notably, the AI outperformed NT-proBNP in head-to-head comparisons and identified HFpEF even in patients with normal biomarker levels. The system requires no new medical hardware, working entirely from standard 12-lead ECG voltage data already routinely collected in clinical settings.