Federated learning of medical concepts embedding using BEHRT.

Objectives: Electronic health record data is often considered sensitive medical information. Therefore, the EHR data from different medical centers often cannot be shared, making it difficult to create prediction models using multicenter EHR data, which is essential for such models' robustness and generalizability. Federated learning (FL) is an algorithmic approach that allows learning a shared model using data in multiple locations without the need to store all data in a single central place. Our study aims to evaluate an FL approach using the BEHRT model for predictive tasks on EHR data, focusing on next visit prediction. Materials and Methods: We propose an FL approach for learning medical concepts embedding. This pretrained model can be used for fine-tuning for specific downstream tasks. Our approach is based on an embedding model like BEHRT, a deep neural sequence transduction model for EHR. We train using FL, both the masked language modeling (MLM) and the next visit downstream model. Results: We demonstrate our approach on the MIMIC-IV dataset. We compare the performance of a model trained with FL to one trained on centralized data, observing a difference in average precision ranging from 0% to 3% (absolute), depending on the length of the patients' visit history. Moreover, our approach improves average precision by 4%-10% (absolute) compared to local models. In addition, we show the importance of the usage of pretrained MLM for the next visit diagnoses prediction task. Discussion and Conclusion: We find that our FL approach reaches very close to the performance of a centralized model, and it outperforms local models in terms of average precision. We also show that pretrained MLM improves the model's average precision performance in the next visit diagnoses prediction task, compared to an MLM without pretraining.