Residual-attention deep learning model for atrial fibrillation detection from Holter recordings

Abstract

Background

Detecting subtle patterns of atrial fibrillation (AF) and irregularities in Holter recordings is intricate and unscalable if done manually. Artificial intelligence-based techniques can be beneficial. In fact, with the rapid advancement of AI, deep learning (DL) demonstrated the capability to identify AF from ECGs with significant performance. However, further development and validation on larger cohorts is still needed.

Purpose

The main purpose of this study was to develop a Residual-attention DL model by considering a large cohort of 2‑lead Holter recordings.

Methods

We developed a residual DL model by collecting a large dataset of 661 Holter recordings, which was labeled manually by an expert cardiologist. The DL model leveraged attention mechanisms, allowing it to capture long-range dependencies and intricate temporal relationships crucial for identifying subtle patterns indicative of AF.

Results

Experimental results demonstrated that our model achieved a sensitivity (detection of AF) of $\mathrm{Se}=0.928$ and a specificity of $\mathrm{Sp}=0.915$, with an AUC-ROC of $\mathrm{AUC}=0.967$ on our dataset. Additionally, when evaluated with an external test dataset, specifically IRIDIA-AF, our DL model obtained $\mathrm{Se}=0.942$, $\mathrm{Sp}=0.932$, and $\mathrm{AUC}=0.965$. Finally, when compared under similar experimental conditions with other state-of-the-art models, our DL model achieved slightly better performance overall.

Conclusion

The Residual-attention DL model we proposed offers a promising solution for AF detection. The validation on external datasets contributes to its potential for deployment in clinical settings, providing clinicians with a valuable decision support system.