Machine learning-based atrial fibrillation detection and onset prediction using QT-dynamicity.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Atrial fibrillation (AF) prediction remains a clinical challenge.
• Ventricular repolarization dynamics, influenced by the autonomic nervous system, may offer predictive insights.
• Existing methods often rely on heart rate variability, potentially missing crucial repolarization information.

Purpose of the Study:

• To evaluate the efficacy of QT dynamicity in predicting paroxysmal AF onset.
• To compare the predictive value of QT dynamicity against traditional ECG features for AF detection and forecasting.
• To utilize interpretable machine learning for analyzing ECG signals and identifying key predictors of AF.

Main Methods:

• Gradient-boosted decision trees (GBDT) were employed for AF prediction using ECG data from 88 patients.
• Wavelet-based signal processing delineated ECG signals, extracting 44 features including QT and RR intervals.
• A patient-level data split (80% train, 20% test) and 5-fold cross-validation ensured robust model evaluation.

Main Results:

• For AF detection, the GBDT model achieved an AUROC of 0.99 and 95% accuracy using a 30s window, with RR interval features being most influential.
• For AF onset forecasting, a 120s window yielded an AUROC of 0.739 and 74% accuracy.
• R wave amplitude and QT dynamicity (QT-RR slope correlation) emerged as the strongest predictors for AF onset.

Conclusions:

• QT dynamicity provides valuable information for accurate short-term prediction of AF onset.
• Ventricular repolarization analysis, specifically QT dynamicity, enhances AF forecasting beyond traditional RR interval and heart rate variability metrics.
• Autonomic nervous system-mediated changes in ventricular repolarization are implicated in AF initiation, highlighting a potential therapeutic target.