Data-Driven Methods for Blood Supply Chain Management Using Electronic Health Record (EHR) Data

运筹学（OR）和机器学习（ML）通常被视为数据驱动决策的两种截然不同的替代方法：运筹学包含大量基于分布假设的决策优化方法，而机器学习提供使用真实和复杂数据的解决方案，然而，通常缺乏可解释性。生物统计学是健康数据科学的基础，包括研究设计、数据收集和结果解释。集成ML、OR和生物统计学可以在医疗保健中架起健康信息学、决策制定和知识转化的桥梁。血液供应链管理（BSCM）是一个重要且具有挑战性的研究领域：重要是因为它对人类生命的价值；然而，由于其复杂性，包括各种产品/患者特征以及来自医疗保健系统、行业和政府的许多不确定因素，因此具有挑战性。低效率的BSCM不仅会导致高昂的医疗成本，还会对献血者和患者造成伤害。医院血库经常做出过多订购血液的决定，导致负面的滚雪球效应，影响到一系列决策，因为一个人向下移动链。ML、OR和生物统计学的仔细集成是否能为BSCM提供更多可解释和可实施的数据驱动解决方案？我最近在红细胞（RBC）库存管理方面的工作显示了这种方法发展在这个问题领域的潜在影响。我的长期研究计划侧重于利用大型电子健康记录（EHR）数据开发数据驱动的血液库存管理方法。在接下来的五年里，我的研究目标是：1。利用电子病历数据开发多产品多设施数据驱动的生产计划和库存管理决策模型。o为多产品多设施的预测开发新的需求预测模型，考虑使用大规模电子病历数据的分层时间序列，以提高预测准确性。o制定综合数据驱动的决策模型，整合多种血液制品的生产和分销，同时考虑到跨产品的相互作用、血液年龄、交付频率和医疗保健政策。o启动联合学习策略，以便在当地医院实际实施。2. 为涉及时间序列预测和优化的数据驱动决策模型开发新的性能评估方法。o通过交叉验证和样本外方法的不同变体，使用现有的性能指标评估数据驱动的决策模型。o为需要定期再培训和验证的数据驱动决策模型开发新的自适应绩效评估方法。这项研究很重要，因为它将导致高质量的集成方法，具有在现实世界实现的巨大潜力。加拿大人将受益于由此产生的数据驱动的决策模型和评估方法，以提高BSCM的效率。

Operations research (OR) and machine learning (ML) are often viewed as two distinct and alternative approaches for data-driven decision making: OR contains a wealth of methods for decision optimization based on distributional assumptions, whereas ML provides solutions using real and complex data, however, often lacks interpretability. Biostatistics are the foundation for health data science, encompassing study design, data collection, and result interpretation. Integrating ML, OR, and biostatistics can bridge health informatics, decision making, and knowledge translation in healthcare. Blood supply chain management (BSCM) is an important and challenging research area: important because of its value for human life; yet challenging due to its complexity including various product/patient characteristics and many uncertain factors from healthcare systems, industries, and governments. Inefficient BSCM not only leads to high healthcare costs, but also results in harm for blood donors and patients. Frequent over-ordering decisions for blood are often made at hospital blood banks, leading to a negative snowball effect impacting a series of decisions as one moves down the chain. Can careful integrations of ML, OR, and biostatistics offer more interpretable and implementable data-driven solutions to BSCM? My recent work on red blood cell (RBC) inventory management shows the potential impacts of such methodology developments in this problem domain. My long-term research program focuses on developing data-driven methods for blood inventory management using large electronic health record (EHR) data. Over the next five years, my research objectives are to: 1. Develop multi-product multi-facility data-driven decision models for production planning and inventory management using EHR data. o Develop new demand forecasting models for multi-product multi-facility predictions considering hierarchical time series using large scale EHR data to improve forecast accuracy. o Formulate integrated data-driven decision models that consolidate the production and distribution of multiple blood products while accounting for cross-product interactions, blood age, delivery frequency, and healthcare polices. o Initiate federated learning strategies for real-world implementation at local hospitals. 2. Develop new performance evaluation methods for data-driven decision models involving time series forecasting and optimization. o Evaluate data-driven decision models using existing performance metrics through different variants of cross-validation and out-of-sample approaches. o Develop new adaptive performance evaluation methods for data-driven decision models that require periodical retraining and validation. This research is important because it will lead to high quality integrated methodologies with great potential for real-world implementation. Canadians will benefit from the resulting data-driven decision models and evaluation methods to improve BSCM efficiency.