Grant: Project Grant - PA: Computer modelling/simulation to accelerate health research

Electronic Medical Records are secure, private, lifetime records containing patient-health and point-of-care histories within the healthcare system, which can be used for disease surveillance, clinical studies, and many others. It is hard to share and use EMR data across jurisdictions because of the differences and barriers in data governance and architecture, as well as data security regulations. Federated learning is a new machine learning paradigm that trains machine learning models collaboratively without centralizing the data or exchanging data among clients, such as hospitals. Federated learning can broaden data used for analysis while honouring the required ethical and data sharing agreements. Although sepsis is a major worldwide health problem, identifying sepsis remains challenging. The definitions for sepsis have been changing over the past 25 years. There is no gold standard definition for sepsis. Wide variation in incidence rates results from sepsis identification beginning with a pre-determined clinical definition and searching for cases in administrative and/or EMR data. Published work primarily utilize limited site-specific administrative data or registry data collected specifically for sepsis case-finding, presenting significant limitations on method generalization. We propose to develop a novel Federated Learning framework for disease surveillance, with applications to the detection of sepsis, using electronic medical record data across multiple provinces in Canada without data sharing. We will evaluate the performance of the proposed framework through comparing with manual chart review, traditional locally trained machine learning models, as well as the commonly used coded definition of sepsis. The developed Federated Learning framework based on electronic medical records can be used to routinely detect sepsis for disease surveillance and monitoring and facilitate decision-making.