Deep learning-based classification using 12-lead electrocardiogram (ECG) exams has been shown to learn and accurately recognize a wide range of heart conditions, such as atrial fibrillation (AF). There is a need for a decentralized, privacy-oriented approach to conduct model training for automated ECG classification without having to share sensitive patient data in a central location, and for ensuring compliance to data privacy regulations in present day healthcare. Federated learning comes forward as a model training scheme for privacy-oriented, collaborative learning for a wide range of learning tasks. However, its efficacy can be challenged by heterogeneous (non-IID) data, and with communication constraints such as clients dropping from update rounds. The CODE15% dataset by the Clinical Outcomes in Digital Electrocardiology (CODE) study from the Telehealth Network of Minas Gerais, Brazil, holds a large collection of labelled digital ECG exams obtained between 2010 and 2017. This thesis explores the training of an AF classifier residual neural network on a subset of the CODE15% dataset, in centralized to federated learning. Its performance is observed under IID and non-IID datasets with varying feature distributions by patient age. Simulations are done with different client participation ratios, and with 3 federated aggregation strategies. Obtained findings show that federated training can successfully match and improve the performance to centralized training of the AF classification model. However, it suffers in reduced participation, and aggregation strategies FedProx and SCAFFOLD aid in reducing deviations to stabilize convergence in non-IID cases, however also lead to increased training times and convergence durations.