Predictive modeling to create a proactive approach to patient blood management in the oncology population.

Abstract

332 Background: Patient safety concerns that arose during COVID-19, related to blood shortages at a large oncological transfusion center, foregrounded the need for predictive modeling tools to optimize blood product inventory control. A maximum surgical blood ordering schedule (MSBOS), is a tool used to assist clinicians in predicting intraoperative blood usage based on retrospective historical data within an institution. Although MSBOS proves to be valuable, it is rudimentary in nature. Not only is data collection cumbersome but the data generated may not reflect current surgical practices and inter-patient variability may skew procedural averages. Predictive blood modeling is contingent generation of a digital health dashboard (DHB). DHB are electronically embedded in the electronic health record (EHR) to collect perioperative data. Coupling the generated informatics (patient demographics, diagnosis, laboratory results, procedural type, medications/supplements, surgeon) with machine learning allows for creation of patient-centered predictive blood modeling algorithms and better inventory control. Methods: To characterize blood use across various procedures at our institution, we engaged information technology specialists to create a Web Intelligence report by integrating data from both an EHR and a lab information system (LIS) into a single repository. Information obtained illuminated a master procedure list, blood product usage patterns, and characterized patient demographics during January - March, 2020. Data is continuously extracted to create a perpetually updated MSBOS while secondarily functioning to cultivate data for future predictive machine learning algorithms. Results: Data analysis demonstrated 5598 procedures were performed during the first quarter of 2020. Procedures not transfused with packed red blood cells (pRBCs) totaled to 4,156 and 1,442 had a greater than or equal to 10% probability of requiring pRBCs. Our current practices reflected our overall crossmatch to transfusion ratio ( C:T) was 5.4 to 1. Concerted collaboration, resulting in preparation of pre-surgical blood product orders according laboratory generated MSBOS schedule could decrease the C:T to 1.7 to 1. Additionally, high intraprocedural pRBCs variability was identified in current procedural subtypes. Conclusions: Traditionally generated MSBOS are functionally limited and may not be reflective of current surgical practices. Additionally, inter-patient variability may distort some procedural type guidance. Creating an integrated data report, eliminates some of the inherent limitations of traditional MSBOS. Moving forward, the cultivated data if coupled with machine learning has the potential to create transferable proprietary algorithms that proactively predict individual patient transfusion needs.