Heart Failure Detection in Electrocardiograms Using Artificial Intelligence

The diagnosis of heart failure (HF) is resource-intensive, requiring specialized personnel and equipment, which often leads to severe underdiagnosis. This study proposes the use of a machine learning model to detect HF in Electrocardiograms (ECGs), a commonly used tool in most healthcare settings. HF is known to have limited validity in diagnosis codes, limiting the viability of direct application of supervised learning. However, we hypothesize that validating diagnosis with measured levels of circulating N-terminal proB-type natriuretic peptide (NT-proBNP), ameliorates the impact of label noise in a large dataset. We demonstrate the success of the labelling strategy by developing a neural network and prospectively validating it in a cohort comprising 43109 patients. The model significantly outperformed NT-proBNP in diagnostic accuracy (p = 7.5e-7), and is capable of detecting both HF with reduced ejection fraction (AUC=0.91) and preserved ejection fraction (AUC=0.68-0.89 depending on definition). To highlight the impact of underdiagnosis in evaluating the model, we conducted a small-scale retrospective clinical evaluation of the test set, including patients with ejection fraction >50% with no HF diagnosis and normal levels of NT-proBNP. In this subgroup, 24 out of the 30 patients with the highest model-predicted risk satisfied the diagnostic criteria for HFpEF. These combined findings demonstrate the model’s capability in finding HF independent of ejection fraction, and a potential for accessible diagnostics through AI-enhanced ECG analysis.