ECG-only explainable deep learning algorithm predicts the risk for malignant ventricular arrhythmia in phospholamban...

Area of Science:

• Cardiology
• Artificial Intelligence
• Genetics

Background:

• Phospholamban (PLN) p.(Arg14del) variant carriers face a high risk of malignant ventricular arrhythmia (MVA).
• Accurate risk stratification is crucial for timely intervention, including intracardiac defibrillator implantation.

Purpose of the Study:

• To evaluate an explainable deep learning (DL) approach for MVA risk prediction using only electrocardiogram (ECG) data.
• To compare the DL model's performance against current multimodality prediction models.

Main Methods:

• A DL-based variational auto-encoder was trained on 1.1 million ECGs to generate a compressed ECG representation (FactorECG) with 32 explainable factors.
• Cox regression models were developed using FactorECG data from 679 PLN p.(Arg14del) carriers without baseline MVA.

Main Results:

• The DL ECG-only model achieved a C statistic of 0.79, comparable to the current model (0.83) and superior to conventional ECG parameters (0.65).
• A 2-step screening approach using the DL model reduced additional diagnostics by 60% while outperforming the multimodal model.
• An interactive visualization tool was developed for clinical application.

Conclusions:

• An explainable DL algorithm utilizing solely ECG data can accurately predict MVA in PLN p.(Arg14del) carriers.
• This approach facilitates more efficient patient stratification for diagnostic testing and follow-up.