Sequential Factorized Autoencoder for Localizing the Origin of Ventricular Activation From 12-Lead Electrocardiograms

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Population analysis of electrocardiograms (ECGs) is challenged by significant inter-subject variations.
• Accurate localization of ventricular tachycardia (VT) origin is crucial for effective treatment planning.
• Existing methods for VT origin localization often struggle to account for individual patient differences.

Purpose of the Study:

• To develop and validate a novel approach for localizing the origin of VT from 12-lead ECGs.
• To specifically address and disentangle inter-subject variations in ECG data during population analysis.
• To improve the accuracy of VT origin localization by isolating patient-specific factors.

Main Methods:

• A factor disentangling sequential autoencoder (f-SAE) was developed, utilizing both LSTM and GRU networks.
• A pair-wise contrastive loss function was introduced to facilitate the disentanglement of inter-subject variations.
• The f-SAE model was trained and evaluated on a large ECG dataset from VT patients with known pacing sites.

Main Results:

• The f-SAE approach demonstrated improved classification accuracy for VT origin by up to 8.94% compared to QRS features and 1.5% over a deep CNN.
• For predicting VT origin coordinates, the f-SAE achieved a superior accuracy, reducing error by 2.25 mm compared to QRS features.
• The f-SAE outperformed a standard sequential autoencoder (SAE) by 5.15% in classification and 1.6 mm in coordinate prediction, highlighting the benefit of factor disentanglement.

Conclusions:

• The factor disentangling sequential autoencoder (f-SAE) is a feasible and effective method for localizing VT origins from 12-lead ECGs.
• The approach successfully separates inter-subject variations, leading to more robust and accurate population-level ECG analysis.
• This work highlights a promising research direction for overcoming inter-subject variability challenges in ECG signal analysis.