Deep learning algorithm for detection of acute heart failure using standard ECG waveforms.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Acute heart failure (HF) diagnosis requires timely and accurate assessment.
• Electrocardiogram (ECG) is a readily available diagnostic tool.
• Deep learning offers potential for advanced waveform analysis.

Purpose of the Study:

• To develop and evaluate deep learning models for acute HF diagnosis using 12-lead ECG.
• To assess model performance on a large internal cohort and external validation dataset.
• To explore the utility of AI in identifying high-risk cardiovascular patients.

Main Methods:

• Retrospective analysis of ECG, echocardiogram, and NT-proBNP data from 1949 acute HF patients and 24,603 controls.
• Development of four deep learning models (1D-CNN-Res, 1D-CNN-Dense, CRT-Net, CRT-Net without transformer) and an ensemble model.
• External validation using the MIMIC-IV dataset (7868 acute HF patients, 16,025 controls).

Main Results:

• The ensemble deep learning model achieved an AUROC of 0.997 and F1-score of 0.649 on the internal cohort.
• External validation demonstrated an AUROC of 0.842 and F1-score of 0.640.
• The model showed consistent performance across diverse ejection fraction values and demographic subgroups.

Conclusions:

• Deep learning models show significant potential for accurate and immediate acute HF diagnosis from ECG.
• The developed AI tool can aid in early detection, potentially improving clinical outcomes.
• Post-hoc analysis suggests utility in identifying patients with underlying cardiovascular risks.