AABB Foundation Grant Recipient Is Working To Develop Machine-Learning Tools to Create Models That Could Identify Personalized IDIs

“AABB Foundation: Shaping the Industry” is an AABB News series that highlights advances in scientific and medical research supported by the AABB Foundation. Give to the AABB Foundation today to help provide critical funding for the blood and biotherapies field.

Tailored risk-based inter-donation intervals (IDIs) could boost blood supplies by allowing more frequent donations by reducing the wait between donations for individuals with quicker iron-recovery, while better protecting donors who require longer periods to replenish iron stores. AABB Foundation 2022 Early-Career Grant recipient W. Alton Russell, MS, PhD, is working to develop machine-learning tools to create models that could identify personalized IDIs. Russell is an assistant professor at McGill University in Montreal.

“If you have a history of higher donation intensity in the last couple years, then you might need more time [to recover iron levels]. If you have a history of deferrals, then maybe you would need more time. Using longitudinal data from the donors allows us to look backwards to see patterns that can inform future risk,” Russell said.

Optimizing Donor Safety

Donors are the foundation of blood-based treatments that save and improve thousands of lives each year, and much research has been dedicated to keeping them healthy and safe over the years, allowing them to be repeat donors. Maintaining iron stores has been a key component of donor health research, particularly for repeat donors. Several options have been — and continue to be — investigated.

These potential solutions include later providing donors with information about their iron stores, measured in blood ferritin levels. Because ferritin levels can’t be measured at the point of donation, hemoglobin continues to serve as a surrogate at the time of donation, whether a donor has adequate iron to prevent complications following donation. This approach has potential consequences for donations, as providing ferritin levels is a two-step process and may dissuade donors from returning.

An alternative approach, minimum donor IDIs require donors to wait a defined period of time before they are eligible to donate again, giving their bodies time to replenish iron stores. Donor deferrals are mostly a one-size-fits-all approach. Some donors rebuild iron stores more quickly than others, and conversely, others may require more time. Blanket deferral periods not only put some donors at risk, but can also decrease the pool of potential donors for an unnecessary period of time.

Russell and his team are developing risk-based machine-learning models that could lead to more personalized deferral periods for donors. Machinelearning models are already used elsewhere in health care to personalize care based on predicted risk where large numbers of patient-level datasets are available.

Russell has used such models to produce individual risk trajectories — estimates of how risk of hemoglobin deferral or giving blood with low or absent iron stores evolve based on the donor’s IDI – using some available datasets.

 Variation in Iron Recovery

Already Russell’s team has identified considerable variation in iron recovery periods, meaning that minimum IDI tailored to each donor’s predicted risk could protect high-risk donors without limiting more frequent donations from low-risk donors.

Russell’s AABB Foundation research award, “Tailoring blood donation intervals to individual risk in South Africa, the Netherlands, and the United States,” will use operational data from South Africa, the Netherlands and the U.S. to develop and analyze new policies that tailor the minimum IDI to each blood donor’s estimated risk of deferral or low/absent iron.

“With this grant we will be able to externally validate the models that we develop based on U.S. in operational data,” he said.

One member of Russell’s team, first-year PhD student Chen-Yang Su, is working on externally validating the machine-learning model already developed using operational data from South Africa, the Netherlands and the U.S.

For another component of the grant project, first-year MSc student Huzbah Jagirdar is working to understand differences in donor return behavior after a completed donation or received a deferral, with a particular focus on how donor return behavior at mobile collection sites differs from those at a fixed collection center.

“The return dynamics of mobile donors is important because they represent a large number of donors in some countries,” Russell said. “We also plan to look at whether it’s possible to develop simpler decision rules that approximate the machine learning model,” he said.

Not all collection facilities would have the ability or capital to use machine-learning models more tailored to their donor population, while others may be uncomfortable with a machine-learning “black box” approach.

“Once we finish the external validation of the machine-learning model, we want to look at whether simpler decision rules can be developed that can provide the majority of the benefits of the machine learning model, but without requiring the same technical skills or resources,” Russell said.

A number of donor behavior questions will also need to be investigated, such as how donors will feel about the minimum IDI not always being the same for them or for friends who have a different interval.

“Ultimately, we can explain to donors that iron recovery is different for everyone … and that donors will understand. But I do think that some research needs to be done in terms of acceptability of risk-based among donors,” Russell said.

For the future, the team plans to study the implementation of their data-driven tailored minimum IDI policies with collaborating blood agencies.