– By Dr Joseph Farrington, winner of the Operational Research Society’s Doctoral Award
Hospital blood banks play a critical role in ordering, storing, processing, managing and issuing blood components, including platelets, which are vital for preventing and treating bleeding in numerous clinical settings. For example, platelets are essential for minimizing blood loss during chemotherapy, with NHS Blood and Transplant reporting that around 69% of donations go to patients with cancer. However, platelets are highly fragile and have a very short shelf life of just three to five days. This creates a daily challenge for blood banks: ordering too many risks wastage of valuable donations, while ordering too few can delay patient care.
During my research in hospitals, I observed that 8–10% of issued platelets were returned unused from wards. By the time these units were returned, many were close to expiry and had to be discarded, highlighting the difficult balance between shortages and waste that blood banks face.
Managing platelet supply is not simply a stock control issue. Demand can change quickly depending on clinical need, plus there is an ethical responsibility to minimise the waste of donated blood products. These pressures make platelet inventory management a particularly challenging operational problem in healthcare.
To help find a solution, my PhD research explored how operational research (OR) and AI could be used to enable hospitals to make better decisions about how platelets are ordered and allocated. By applying mathematical modelling and machine learning techniques, I found that it is possible to improve supply, reduce waste and support clinicians in delivering safer, more efficient care.
Using accelerated optimisation for smarter ordering
One focus of my research was stock ordering. Each choice about how many platelets to order affects stock levels, shortages, and wastage in the following days. In theory, an optimal strategy can be calculated by modelling these effects, but the number of possible stock combinations grows rapidly, making computation slow and impractical.
To overcome this, I used graphics processing units (GPUs), the hardware behind modern AI systems. Using the open-source library JAX, I could perform many calculations simultaneously, solving problems that once took days in just hours on a desktop GPU. This allowed analysis of realistic supply models and comparison of standard practice against optimal strategies.
Reducing waste by predicting returns
Ordering however, is only part of the challenge. Blood banks must also decide which unit to issue. The standard approach is to issue the oldest unit first. While usually effective, this can increase waste if a unit is likely to be returned.
I explored an alternative issuing policy: “youngest unit for predicted returns.” The principle is simple. If a request is likely to lead to transfusion, issue the oldest unit; if it is likely to be returned, issue the freshest unit, giving it more shelf life if it comes back.
Before developing a complex predictive model, I first asked a more basic question: even with a highly accurate predictive model, would this policy meaningfully improve outcomes? Accessing sensitive data, validating algorithms, and implementing systems requires time and resources, so it is important to establish impact first.
I built a simulation of the blood bank’s daily operations, modelling predictive systems with varying accuracy and measuring effects on wastage and service levels. The results were encouraging. The youngest-unit policy reduced platelet wastage by around 14 per cent while maintaining high service levels.
This simulation-first approach showed how predictive models can be evaluated for real-world value before significant investment. In healthcare, where data access and implementation are complex this allows organisations to prioritise models that can deliver tangible operational benefits rather than accuracy alone.
Multiple uses beyond blood banks
Advances in AI hardware and software make large-scale optimisation practical, while OR provides the systems perspective needed to judge models by operational outcomes rather than statistics alone. Although this work focused on platelet management at one UK hospital, the principles extend beyond blood banks.
Many healthcare resources, from medicines to vaccines, are perishable and face similar tensions between waste and shortage. With the right analytical tools and focus on operational impact, hospitals can move beyond reactive decision-making to smarter, data-driven systems, making better use of donated blood, reducing avoidable waste, and ultimately improving patient care.