August 29, 2024
This article originally appeared in AABB News, a benefit of AABB membership. Join AABB today to read the rest of this month’s issue.
In the past few years the term artificial intelligence (AI) seemed to rise from the pages of science-fiction and become part of the everyday vernacular. Rather suddenly, many everyday interactions with technology are offering opportunities to engage with AI. However, for more than a decade, many in data sciences and health care have recognized the potential of AI to play an important role in harnessing “big data,” streamlining processes and improving patient care.
In fact, one of the key research areas identified at the National Heart, Lung, and Blood Institute’s 2022 State of the Science (SoS) in Transfusion Medicine Symposium was “new computational methods in transfusion science, such as artificial intelligence and machine learning.”
Many in the field have already been working hard for years developing and testing potential uses for AI in transfusion medicine and blood banking.
“When delineating the trajectory of the next decade’s research in transfusion medicine, prioritizing the integration of AI and Machine Learning tools emerges as a pivotal step,” said Ruchika Goel, MD, MPH, CABP, senior medical director, corporate medical affairs at Vitalant, and professor of Internal Medicine and Pediatrics and Southern Illinois University School of Medicine, who co-chaired the data sciences subsection of the SoS symposium. “There is a need to incorporate these tools as much as possible in the fundamental elements of clinical practice, research and education.”
The possible applications of AI in transfusion medicine are myriad. One potential use involves harnessing technology for better blood utilization predictions and calculations.
Kelley Counts, MS, MBA, director of data science at OneBlood; and Karl Rexer, PhD, founder of Rexer Analytics, have been successfully forecasting overall (cross-hospital) blood order volume for years, and it enabled OneBlood to right-size blood collection efforts throughout the COVID-19 pandemic. Counts and Rexer outlined that during the pandemic, hospital blood orders dropped 30-40%.
“Our monitoring and forecasting system enabled us to respond quickly,” Counts said. “We reduced blood collections to match the declining orders. Many donors and staff could be safely mitigated from frontline collections and were not put at risk during the pandemic, and when hospital orders rebounded, our system immediately saw it, and we gradually ramped collections back up.”
Counts and Rexer have also begun work to develop anticipatory ordering.
“This concept is something that companies like Amazon use a lot,” Counts explained. “They want to anticipate when a customer will run out of something so that when they do, the new item will land on their doorstep and customers are never without the product they need.”
Counts emphasized that their work is still exploratory and not yet deployed, but that back-testing on historical data looks very promising. He and Rexer looked at the use of anticipatory ordering applied to O-positive blood, which accounts for a significant portion of blood product orders and shipments. The goal was to prevent critical hospital inventory shortages. To do that, they developed an algorithmic system that calculated optimal inventory, and recommended the best daily blood order for each hospital.
“We didn’t want to simply ship the average order because with blood ordering, using the average is just about never right,” Counts said.
Instead of focusing on past orders, they focused on blood usage and inventory levels, and they examined both at the average and variance. This helped them understand optimal inventory levels.
“In some simulated cases blood was used like a sewing machine – all the time, but others ebb and flow with usage going up and down,” Counts said.
As the average changed, the variance typically changes, and the optimal inventory can change as well. If a hospital transfused 10 units today and the optimal inventory didn’t change, then the system would send 10 units tomorrow. If the hospital used 10 units and optimal inventory went up by two units, then the system would send 10 units that were used plus an additional two units to meet optimal inventory.
“We have simulated our anticipatory ordering system across a year’s worth of data and the system does a great job adapting,” Counts said. “We haven’t seen instances of critical inventory shortages.”
In addition, with the anticipatory system, the order sizes were lower compared to current manual ordering: the anticipatory system submitted a higher cadence of smaller orders.
“A key benefit we see in our simulations is that with anticipatory ordering, fewer discards may be possible,” Rexer said. “You don’t need to overstock blood, but the system makes sure they never run out of it; it allows them to use blood more efficiently.”
Na Li, PhD, of University of Calgary, Canada, attempted to integrate machine learning and statistical modeling to optimize red blood cell inventory. In Canada, different electronic data collection systems are used across hospitals to collect patient records, but they are not used for inventory management solutions. Li and colleagues wanted to use this data to improve blood demand and supply management. With their forecasting and inventory management strategy, the predicted daily product demand accurately reflected the actual demand; however, the model’s proposed daily ordering quantity was significantly lower than the actual blood bank orders.
“Using a hybrid computation approach, we can get a lean inventory in hospital blood banking, and reduce the cost involved,” Li said. Without risk of shortages, the model’s ordering strategy reduced inventory levels by almost 40% leading to a cost reduction of 43%.
In a review of studies of strategies that use novel computational techniques for blood demand forecasting, Li and colleagues discussed some of the current limitations to these approaches, including generalizability of a study of one blood component to other blood components, lack of ABO compatibility consideration, and some ethical challenges that exist related to resource allocation.
One of the primary ethical challenges in blood demand forecasting is the equitable allocation of limited blood resources.
“Advanced computational techniques can help with optimizing the prediction and distribution of blood components, but they must also ensure that these processes do not inadvertently exacerbate existing inequalities,” Li said. “For example, algorithms may need to be carefully designed to eliminate bias that could lead to preferential treatment of certain populations over others.”
Furthermore, there are ethical considerations regarding the transparency and accountability of these computational models, as well as the need to balance efficiency with fairness in the allocation of such life-saving resources.”
Blood inventory management is only one area for possible application of these technologies. Another example discussed by Goel was the use of a publicly available AI system to classify post-transfusion adverse events.4 Researchers entered multiple transfusion reaction scenarios into the generative AI program ChatGPT-3.5 and asked it to identify the case definition criteria, severity and imputability for a suspected transfusion reaction scenario using the National Healthcare Safety Network criteria. Responses were reviewed and compared with the classification assigned by transfusion medicine specialists.
The AI was able to accurately classify some reaction types, such as transfusion-associated circulatory overload and transfusion-related acute lung injury cases. However, AI was not able to identify certain reactions like cases of delayed serologic transfusion reactions. Overall, its accuracy was below that of the specialists.
“It was very promising that AI-based large language model [LLM] could identify and classify many transfusion reactions with a fair degree of accuracy,” Goel said. “But there were several key limitations, as it could not comment on the reaction severity and imputability indicating that this technology could potentially assist an expert, but it is not meant to replace clinical expertise.”
AI-driven predictive modeling could also be used to determine whether blood products will be needed by patients, such as patients with bleeding admitted to the intensive care unit (ICU), for example, Goel said. Using a machine learning algorithm trained on two publicly available ICU databases, researchers attempted to see if AI could predict ICU transfusion need. The data included more than 10,000 patients admitted to the ICU with gastrointestinal hemorrhage. The model looked at data from the first five hours of ICU admission to predict the need for transfusion in the next 24 hours, and achieved an area under the receiver operating characteristic curve of more than 0.80.
Goel summarized that as per the authors, the algorithms (while preliminary for now), if validated prospectively, could ultimately help to inform clinical decision making. She went on to add in combination with researchers at University of Calgary and Johns Hopkins, her research team is developing machine learning models to explore associations with transfusions and some clinical outcomes like venous thrombo-embolism.
Outside of these limited examples, AI methods could be applied even more widely in transfusion medicine to areas of donor recruitment or retention, predicting transfusion outcomes and in transfusion education. Most of these algorithms and ideas will fail to have an impact though, if they are not successfully deployed and adopted, Goel said.
West Virginia University rolled out an AI-based blood utilization calculator that ran a proprietary AI algorithm to recommend the number of packed red blood cells required to achieve a target hemoglobin or hematocrit value. During the testing phase, based on this AI-based calculator, the target hemoglobin value was achieved in more than 96% of prescribed transfusions.6 However, the tool was significantly underutilized in real-world setting. A qualitative study exploring the underuse of the calculator showed that “theoretical efficacy alone does not ensure technology use or acceptance”; things like system design, end-user perception and users’ knowledge of the technology must also be addressed.
“The real-time implementation and uptake [of the blood utilization calculator] did not go as expected,” Goel said. “This shows that even when effective technology may be available to an end-user, the actual implementation and acceptance often times remains low.”
End-user adoption is a well-known issue in the AI community. “Gartner has routinely reported that if you start 100 data science projects, 15 make it to deployment; 85 fail,” Counts said. “That is routine; if you tip over 25 rocks, you will find only a couple that have something that is valuable and deployable.”
Rexer Analytics has done surveys of analytics professionals since 2007 and has found that the number of projects that get deployed is surprisingly small.
Indeed, Li said that ease of use is a big concern among her colleagues in transfusion medicine.
“We have to increase efforts to test [these tools] and increase customization within the system to make it easier to use,” Li said. “I am working with people within health services on a project to implement a different AI tool into the Epic system and that involves learning the entire journey from the technical side of how to build the interface, to how to integrate it with Epic, to testing and validation of its usability.”
To address barriers, Goel said that developers have to be sure there is effective integration of an AI-based support system. Any project must have multidisciplinary engagement and employ good initial and ongoing training of the end users.
“We must also ensure that these tools evolve from research initiatives and, once validated, become adaptable and deployable in routine practice,” Goel said. “For this, they have to be clinical grade, generalizable and scalable."
“It can be very intimidating to suddenly be expected to be aware of the new technological developments and to incorporate it into our routine workflows,” Goel added. “Recognizing that AI is a novel concept and technology for everyone, striving to remain abreast and gaining a head start in its understanding and integration is paramount.”
Goel emphasized attempting to embrace the technology rather than be intimidated by it. That means asking questions and being inquisitive about how it can help our rapidly transforming field.
“We have to learn how to make AI work for us,” Goel said, “so we can maximize its benefit for our clinical practices and our patients. It is not meant to substitute us, but rather to support us!”
4550 Montgomery Avenue
Suite 700, North Tower
Bethesda, MD 20814
301.907.6977
