Evaluation of Quantitative Decision-Making for Rhythm Management of Atrial Fibrillation Using Tabular Q-Learning

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Rhythm management in atrial fibrillation (AF) presents complex clinical decisions.
• Identifying optimal strategies for individual AF patients remains challenging.
• Clinical trials identify patient subsets but lack personalized guidance.

Purpose of the Study:

• To apply artificial intelligence (AI), specifically tabular Q-learning, to determine optimal initial rhythm-management strategies for atrial fibrillation (AF).
• To compare AI-recommended strategies against provider-selected treatments for patient outcomes.
• To demonstrate dynamic learning capabilities of AI in refining treatment recommendations.

Main Methods:

• Retrospective analysis of 52,547 new-onset atrial fibrillation patients (2010-2020).
• Clustering patients into 8 phenotypes using variational autoencoder and K-means.
• Applying tabular Q-learning with a composite reward function (mortality, treatment change, sustainability).

Main Results:

• Rhythm-control strategies showed superior outcomes across all patient clusters compared to rate-control.
• Patients receiving AI-recommended treatments had significantly lower mortality (8.5% vs. 22.4%) and higher rewards.
• AI demonstrated dynamic learning, updating optimal strategies from cardioversion to ablation in some clusters.

Conclusions:

• Tabular Q-learning offers a dynamic and interpretable AI approach for clinical decision-making in atrial fibrillation.
• AI-guided rhythm management can lead to improved patient mortality and treatment sustainability.
• Prospective validation of Q-learning in clinical practice is warranted.