@phdthesis{discovery10204427,
           month = {February},
           title = {Machine learning and GPU-accelerated computing applied to platelet inventory management in a hospital blood bank},
           pages = {1--432},
            note = {Copyright {\copyright} The Author 2025.  Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/).  Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms.  Access may initially be restricted at the author's request.},
            year = {2025},
          school = {UCL (University College London)},
          author = {Farrington, Joseph Marc},
        abstract = {Hospital blood banks must maintain an adequate stock of platelets and other blood
products to ensure that patients can receive transfusions when needed, while trying
to minimize wastage. Challenges include short product shelf lives, donor-recipient
compatibility, and the fact that some requested units may not be transfused.
I used machine learning (ML) and graphics processing unit (GPU)-accelerated
computing methods to find policies for two key decisions: how many platelet units
to order from a supplier (replenishment) and selecting platelet units to meet demand
(issuing).
My review of the literature identified opportunities to: use deep reinforcement
learning (DRL) to learn how to act in the large state spaces needed to represent
perishable stock; use GPU hardware to compute the optimal replenishment policy
for larger, more realistic problems; and improve issuing policies which have
received less attention than replenishment decisions.
I first developed a novel reinforcement learning environment to model a
simple platelet replenishment problem. DRL policies performed near-optimally on
simulated data and outperformed commonly used heuristic policies on real demand
trajectories from a UK hospital.
My GPU-accelerated implementation of value iteration enabled the optimal
policy to be computed for perishable inventory management problems where this
was recently deemed infeasible or impractical, with up to 16.8M states, using
consumer-grade hardware. These results can support benchmarking approximate
approaches including DRL.
I designed a novel ML-guided issuing policy to address the fact that not all requested platelet units are transfused. I explored how the utility of the policy
depended on the quality of patient-level ML predictions of transfusion and trained
an ML model, with AUROC 0.74, sufficiently good to support wastage reductions
under the new policy.
Finally, I extended the problem to jointly optimize replenishment and issuing
policies to manage multiple perishable products with substitution, including
platelets with all eight ABO/RhD blood types.},
             url = {https://discovery.ucl.ac.uk/id/eprint/10204427/}
}