Using AI to match blood donors with the products hospitals need

Context and objectives

The organization manages the collection and distribution of blood products to hospitals across a region, a process shaped by several converging pressures:

• Fluctuating hospital demand, driven in part by evolving patient care strategies that call for more dynamic stock management

• Rising demand for plasma, growing by more than 5% a year, requiring stronger donor recruitment and retention

• The diversity and rarity of blood groups, making it difficult to guarantee availability across every type, especially the rarest ones

• Donor outreach that was only partially targeted, with no tools to help decide who to contact, when, and for what

At the heart of the challenge was a single, deceptively simple question: which donor should be contacted, for which type of donation, to best match supply with actual demand?

The objective was to explore whether AI could help answer that question directly, by building a proof of concept that predicts, for each donor and each donation type, how likely they are to respond positively to a solicitation. The solution also needed to account for current stock levels and the particular need to protect rare blood types.

Approach

Agilytic and the client's teams worked together to turn this question into a working model, treating it as both a data science challenge and an optimization problem.

1. Understanding the donation process

The project began by digging into how donor solicitation actually worked in practice:

• Which factors influenced a donor's likelihood to respond

• How stock levels for each blood product evolved over time

• Where the current process left the most value on the table

This groundwork mattered as much as the modeling that followed, since predicting donor behavior meant first understanding the real constraints behind each decision to call.

2. Building the prediction models

For each blood product (blood, platelets, plasma), Agilytic built a dedicated model estimating each donor's probability of responding positively to a solicitation. Rather than treating all donors the same way, the models could be retrained over time as new data came in, keeping predictions aligned with how donor behavior actually evolves.

The project also presented a significant challenge from a GDPR and data privacy perspective, given the sensitive nature of donor information. Privacy was therefore built into the approach from the outset: the work relied on anonymized data or privacy-preserving solutions and methods designed to safeguard confidentiality throughout the project.

3. Turning predictions into action

Knowing who is likely to respond isn't enough on its own. Agilytic added a constraint optimization layer on top of the prediction models, generating prioritized calling lists that balanced response likelihood against real operational needs. This included current stock levels, with a focus on keeping rare blood types available even when the odds of a response were lower.

Results

The project delivered a working proof of concept demonstrating how AI could support one of the organization's most delicate balancing acts: keeping the right blood products available without over-soliciting donors.

Key deliverables were:

• A prediction models for each blood product type, estimating individual donor response likelihood

• The underlying code to retrain these models over time, so predictions could keep pace with evolving donor behavior

• A constraint optimization engine turning those predictions into prioritized, actionable calling lists

The approach was validated through field testing, going beyond a purely theoretical proof of concept.

The work was also recognized more broadly, featured in the press and presented at a conference, a reflection of the project's relevance to a challenge many blood services face.