Prognostic Significance and Associations of Neural Network-Derived Electrocardiographic Features

Area of Science:

• Cardiology
• Artificial Intelligence
• Genetics

Background:

• Subtle electrocardiogram (ECG) features crucial for prognosis may be missed by clinicians.
• Supervised machine learning can identify thousands of ECG features beyond conventional parameters.
• The study explores the predictive power of neural network-derived ECG features for future health outcomes.

Purpose of the Study:

• To investigate if neural network-derived ECG features can predict future cardiovascular disease and mortality.
• To identify phenotypic and genotypic associations related to these ECG features.
• To leverage artificial intelligence for enhanced cardiovascular risk assessment.

Main Methods:

• Extracted 5120 neural network-derived ECG features using an AI-enabled ECG model.
• Applied unsupervised machine learning to identify three distinct phenogroups.
• Validated findings across five diverse international cohorts (US, Brazil, UK) with over 1.8 million patients.

Main Results:

• Three identified phenogroups showed significantly different mortality profiles.
• Phenogroup B demonstrated a 20% increase in long-term mortality (HR 1.20) and higher risks of atrial fibrillation, ventricular tachycardia, ischemic heart disease, and cardiomyopathy.
• Genome-wide association studies identified four loci, including SCN10A, SCN5A, CAV1, and ARHGAP24, with roles in cardiac function and potential novel targets.

Conclusions:

• Neural network-derived ECG features effectively predict all-cause mortality and future cardiovascular diseases.
• Biologically plausible and novel phenotypic and genotypic associations were identified.
• These findings offer insights into the mechanisms underlying increased cardiovascular risk detected by AI.