UC San Diego

The broad adoption of Electronic Health Records (EHRs) accelerated the development and usage of Machine learning (ML) and Deep learning (DL) algorithms in clinical settings. The potential uses of ML and DL algorithms to augment clinical decision-making in domains such as forecasting disease onset and progression, predicting response to treatments, and optimization of treatment protocols are growing. While most existing ML/DL models are trained on single-centered data, multi-center datasets are becoming increasingly available. However, curation of such datasets is often time-consuming and lags behind shifts in disease prevalence and changes in workflow practices, which are known to cause data distribution shifts and degradation in ML/DL model performance.%\cite{piccialli_survey_2021}%

In addition, data privacy concerns and patient confidentiality regulations continue to pose a major barrier to multicenter EHR data access. In this work, we developed algorithms to enable DL models to transfer their knowledge across institutional boundaries and learn from new episodes of patient care without forgetting previously learned patterns. We validated and compared our methods in the context of early prediction of sepsis using data across four geographically distinct healthcare systems. We explore several methods to enhance the generalizability of DL models. We focus on three areas: Continual Learning, Federated Learning, and Generative Adversarial Networks (GANs), introducing new algorithms within each area and comparing their performance against state-of-the-art models. We have validated and compared these methods in one of the most challenging tasks for biomedical researchers: predicting the onset of sepsis in intensive care units.