Heart failure detection in electrocardiograms using Artificial Intelligence and pragmatic labelling.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Heart failure (HF) diagnosis is complex and resource-intensive, often leading to underdiagnosis.
• Current diagnostic codes for HF have limited accuracy, hindering supervised machine learning approaches.
• Electrocardiograms (ECGs) offer a non-invasive data source for potential HF detection.

Purpose of the Study:

• To develop and validate a deep learning model for detecting heart failure (HF) using solely electrocardiogram (ECG) data.
• To address the challenge of noisy diagnostic codes by incorporating NT-proBNP levels for improved training data.
• To assess the model's performance across diverse patient cohorts and validate its ability to capture cardiac function.

Main Methods:

• A neural network was developed and trained on a large dataset (25,300 patients) with ECGs and associated diagnostic codes.
• Label noise in diagnostic codes was mitigated by validating with measured N-terminal proB-type natriuretic peptide (NT-proBNP) levels.
• Prospective validation was performed on an independent cohort (43,727 patients) and external validation using the MIMIC-IV database (161,352 patients).

Main Results:

• The model achieved an AUC of 0.86 for hospital-diagnosed HF, increasing to 0.91 and 0.96 with NT-proBNP-adjusted thresholds.
• External validation in the MIMIC-IV cohort showed comparable AUCs of 0.87, 0.90, and 0.96.
• Model predictions correlated well with echocardiographic assessments, indicating accurate capture of both diastolic and systolic cardiac function.

Conclusions:

• Deep learning models utilizing ECGs can effectively detect heart failure, overcoming limitations of traditional diagnostic codes.
• Incorporating NT-proBNP levels significantly enhances the model's diagnostic accuracy.
• The open-source model shows promise for improving HF detection and management, with validated performance across large, independent datasets.