Towards federated transfer learning in electrocardiogram signal analysis

Modern methods in artificial intelligence perform very well on many healthcare datasets, at times outperforming trained doctors. However, many assumptions made in model training are not justifiable in clinical settings. In this work, we propose a method to train classifiers for electrocardiograms, able to deal with data of disparate input dimensions, distributed across different institutions, and able to protect patient privacy. In addition, we propose a simple method for creating federated datasets from any centralized dataset. We use autoencoders in conjunction with federated learning to model a highly heterogeneous modeling problem using the Massachusetts Institute of Technology Beth Israel Hospital Arrhythmia dataset, the Computing in Cardiology 2017 challenge dataset, and the PTB-XL dataset. For an encoding dimension of 1000, our federated classifier achieves an accuracy, precision, recall, and F1 score of 73.0%, 66.6%, 73.0%, and 69.7%, respectively. Our results suggest that dropping commonly made assumptions significantly complicate training and that as a result, estimates of performance of many machine learning models may overestimate performance when adopted for clinical settings.