Cross-province federated machine learning of electronic health records in Canada

加拿大各省电子健康记录（EHR）系统的快速和广泛采用为开展变革性健康信息学研究创造了丰富的数字医疗数据和机会。快速发展的机器学习（ML）技术为疾病预测和监测中的大规模EHR挖掘提供了巨大的希望。然而，EHR数据访问一直是主要障碍，因为EHR数据高度敏感，并且其使用受到严格监管。FL通过协作训练ML模型来规避数据访问问题，而无需在医院之间交换数据。我们提出了一个新的联邦学习（FL）框架，使训练强大的ML模型的真实世界的EHR数据集从加拿大医院跨越加拿大3个省，即魁北克，安大略，和阿尔伯塔。我们将通过3个独立的目标解决FL中的3个计算挑战：目标1：我们将开发一个可部署到加拿大医疗保健系统的强大且隐私保护的FL;目标2：我们将考虑由于来自3家省际医院的异质人群而导致的数据分布变化;目标3：我们将在三个医疗保健部门建立一个联合的自动化人口疾病监测系统。我们提出的项目将大大加快从局部逐块健康信息学研究到全国范围内的医疗保健研究的范式转变。我们提出了一个详细的培训计划，以培养下一代AI+Health研究人员。HQP将由团队成员在他们的研究中心（麦吉尔大学，卡尔加里大学和UBC）进行监督，并进行夏季交流访问。我们将每月举办一次实验室联席会议，以丰富学员的跨学科研究经验。由于外语在学术界和工业界的发展趋势，HQP在完成拟议的健康外语项目时所获得的技能在就业市场上受到高度追捧。

The rapid and broad adoption of electronic health records (EHR) systems across provinces in Canada has created rich digital healthcare data and opportunities for conducting transformative health informatics research. Rapidly advancing machine learning (ML) technologies offer great promise for large-scale EHR mining in disease prediction and surveillance. However, EHR data access has been the main hurdle because EHR data are highly sensitive, and their usage is tightly regulated. FL circumvents the data access problem by training ML models collaboratively without exchanging the data among the hospitals. We propose a novel Federated Learning (FL) framework to enable training robust ML models on real-world EHR datasets from Canadian hospitals spanning across 3 provinces in Canada, namely Quebec, Ontario, and Alberta. We will address 3 computational challenges in FL with 3 separate Aims: Aim 1: We will develop a robust and privacy-preserving FL that is deployable to the Canadian healthcare systems; Aim 2: We will account for the data distribution shifts due to heterogeneous populations from the 3 inter-provincial hospitals; Aim 3: We will develop a federated automated population disease surveillance system across the 3 healthcare sectors. Our proposed project will greatly accelerate the paradigm shift from local piece-wise health informatics research to nation-wide healthcare research. We propose a detailed training plan to train the next-generation AI+Health researchers. The HQP will be supervised by team members at their sites (McGill University, University of Calgary, and UBC) with summer exchange visits. We will host joint monthly meeting among the labs to enrich trainees' interdisciplinary research experience. The skills acquired by the HQP in completing the proposed FL-in-health project are highly sought in the job market due to the trend of FL in both academia and industry.