Atrial fibrillation (AF) is a common heart condition affecting the elderly population and is a significant risk factor for strokes, making it a growing public health concern. Catheter ablation (CA) is the most effective long-term treatment for persistent AF. Recently, a novel CA approach based on spatiotemporal dispersion (STD) has been pro- posed. This technique targets STD patterns associated with active zones responsible for sustaining the arrhythmia. We present three datasets to be used to train and test different machine learning models in automatically identifying STD patterns from multipolar electrograms (EGM). Two different real dataset have been acquired from Nice Pasteur University Hospital and labelled by experts. To address the challenging scenario presented by the real data, a synthetic dataset has been created generating EGM records resembling real-world scenarios, using openCARP cardiac electrophysiology simulation software. We evaluate 13 machine learning techniques to demonstrate the challenging scenario of the real data, and we analyze their performance in the proposed datasets. Results show that the synthetic data are promising as training set for classifiers evaluated on real data, but a deeper statistical analysis is necessary to confirm these findings.