Cardiovascular diseases, particularly Atrial Fibrillation (AF), remain a significant global health burden. Despite advancements in diagnostic and treatment techniques, the long-term success rates of AF ablation procedures remain suboptimal. This is primarily due to the complexity of underlying mechanisms, challenges in accurately identifying arrhythmogenic substrates, and the efficacy heavily relying on physician interpretation, which contributes to variability in procedural outcomes. This work addresses these limitations by leveraging machine learning (ML) for the classification of persistent AF using multi-lead electrograms (EGMs). We investigate logistic regression with handcrafted features, as well as convolutional neural networks (CNNs) and Long Short-Term Memory (LSTM) networks designed to interpret transformed EGMs and capture temporal dependencies, with the aim of enhancing the accuracy of identifying regions suitable for ablation. All three approaches show promise in identifying persistent AF behavior, even in data-limited settings. These results highlight the potential of ML to improve diagnostic precision and support more effective, personalized ablation strategies for persistent AF.