Functional feature extraction and validation from twelve-lead electrocardiograms to identify atrial fibrillation

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Deep learning on electrocardiograms (ECGs) can predict atrial fibrillation (AF) risk but lacks clinical interpretability.
• Existing methods do not elucidate individual-level electrocardiographic changes associated with AF development.

Purpose of the Study:

• To propose and validate a nonparametric feature extraction approach for identifying ECG features linked to incident atrial fibrillation (AF).
• To provide clinicians with interpretable insights into ECG alterations preceding AF onset.

Main Methods:

• Functional principal component analysis (FPCA) applied to raw ECG tracings from the Chronic Renal Insufficiency Cohort (CRIC) study.
• Feature selection using Phase I ECGs (2003-2008); association with incident AF risk evaluated using Cox proportional hazards models.
• Validation of identified features and their longitudinal changes in Phase III ECGs (2013-2015).

Main Results:

• Four ECG features related to P-wave amplitude, QRS complex, and ST segment were identified.
• Both initial measurements and 3-year changes in these features are associated with incident AF risk.
• A 3-year decline in P-wave amplitude (1 SD) independently increased AF risk by 29% (HR: 1.29).

Conclusions:

• The proposed nonparametric features are intuitive and offer interpretability, unlike deep learning models.
• This approach provides insights into individual-level longitudinal ECG changes that precede AF development.
• The findings facilitate a better understanding of AF pathophysiology through ECG analysis.