Federated Learning (FL) enables multiple clinical institutions to jointly train models without sharing raw patient data, but its practical performance relative to centralized approaches remains poorly characterized under realistic, heterogeneous conditions. This study compared isolated, centralized, and federated training paradigms for detecting myocardial infarction (MI) from electrocardiograms (ECG), using PTB-XL and PTB Diagnostic as two structurally distinct simulated clients differing in size, sampling rate, and lead configuration. An InceptionTime architecture was trained identically across paradigms, with FL implemented via Flower using FedProx to limit client drift. On the full-size datasets, isolated and centralized training outperformed FL on the large PTB-XL dataset, while FL outperformed centralized training on the smaller PTB-DB dataset. When PTB-XL was downsampled to match PTB-DB's size (Equal-XL), performance differences between the three training paradigms became statistically insignificant. These results show that FL is particularly advantageous for smaller participants when there are large imbalances in dataset sizes across members of the data-sharing network. The results also indicate that aligning heterogeneous cardiology datasets from multiple sites remains a major practical obstacle to implementing FL, regardless of the specifics of model training.