Site-Aware Federated Learning via Embedding and Resampling with Electrocardiograms

Modern machine learning (ML) methods perform remarkably across a number of diagnostic tasks. Despite this performance, the integration of ML methods in healthcare is relatively limited. While there are a variety of reasons for this, it is notable that most approaches ignore additional constraints that must be made in the healthcare setting. In particular, there may be a relative paucity of data from any single institution; therefore, collaboration is necessary in order to amass a dataset suitable for ML. Furthermore, data may be heterogeneous, with different labels and different input dimensions. Finally, respecting patient privacy is paramount. In this study, we train a classifier under the assumptions of (1) data distributed across multiple institutions, (2) highly heterogeneous data, and (3) a requirement for patient privacy. We enable site-awareness using a global average pooling module to capture high-level information about electrocardiogram (ECG) recording methods combined with a ResNet to encode specific features in ECGs, and we demonstrate that the proposed site-aware ResNet (SA-ResNet) outperforms other state-of-the-art approaches in cardiovascular disease diagnosis. On a highly heterogeneous dataset constructed from three independent datasets distributed unevenly across seven institutions, the proposed model achieves an accuracy, precision, recall, and F1 score of 76.3%, 69.5%, 76.8%, and 73.0%, respectively.