Detection of hospital-admitted heart failure regardless of ejection fraction using artificial intelligence-enhanced electrocardiograms

Abstract Heart failure (HF) presents with diverse and often non-specific symptoms, frequently overlapping with those of other conditions, making it challenging to diagnose accurately. Developing a more affordable and widely available diagnostic method that provides immediate feedback could improve the care and management of patients with HF. Previous work on AI-enabled ECG for detecting HF has largely been based on echocardiographic labelling for model training, and often focused on particular phenotypes of HF like HF with reduced or preserved ejection fraction (EF). Our aim was to develop an AI model identifying HF across the EF spectrum and to show that explicit echocardiographic labelling is in fact not necessary, and that a cheaper data labelling strategy results in AI models with desirable diagnostic properties. ECGs from a network of four tertiary hospitals in the period 2016-2024 were used, with data from 2016-2022 for development (40,000 patients, 161,000 ECGs) and 2023-2024 (43,000 new patients) for prospective testing. Using our proposed labelling strategy based on ICD-10 codes and NT-proBNP levels in the development set, 20,000 ECGs were labelled as HF and 58,000 as non-HF, while 83,000 ECGs were excluded due to uncertain status i.e. NT-proBNP between 125 and 1000 ng/L. An ensemble of cross-validated AI models was evaluated on the first ECG of each of the 43,000 patients (3,500 with HF diagnosis), including 5,000 who had an echocardiogram within 14 days. The model achieved high diagnostic accuracy for hospital-diagnosed heart failure (HF) with an AUC of 0.84 (95% CI 0.831-0.848), regardless of EF and NT-proBNP levels. For the population with measured NT-proBNP levels, the model outperformed NT-proBNP in diagnostic accuracy (p = 7.5e-7). In the test set with echocardiography, the model detected HF with reduced EF with an AUC of 0.91 (95% CI 0.894-0.931) and HF with preserved EF with an AUC of 0.68-0.89, depending on definition. Comparison with the H2FPEF risk score showed that model-predicted risk increased with the score and further stratified risk within each category. To highlight the impact of underdiagnosis in evaluating the model, we conducted a small-scale retrospective clinical evaluation of 60 patients with EF 50%, no HF diagnosis, and normal NT-proBNP levels (125 ng/L). Out of the 60 patients, 30 had a low model-predicted risk while 30 had a high risk; 24 of the 30 highest-risk patients met HFpEF criteria, while 27 of the 30 lowest-risk patients did not. Our study shows that using a dataset of patients with clinical HF or no HF validated via NT-proBNP in the development phase, a neural network can be trained to detect of HF independent of EF, including HFpEF with low NT-proBNP. The model could serve as an accessible diagnostic marker for HF in primary care, as a screening tool to identify patients who may benefit from further evaluation by e.g. echocardiography.Model HF probability by LVEF Model HF probability by H2FPEF