Title:Many happy returns: machine learning to support platelet issuing and waste reduction in hospital blood banks

Abstract:Efforts to reduce platelet wastage in hospital blood banks have focused on ordering policies, but the predominant practice of issuing the oldest unit first may not be optimal when some units are returned unused. We propose a novel, machine learning (ML)-guided issuing policy to increase the likelihood of returned units being reissued before expiration. Our ML model trained to predict returns on 17,297 requests for platelets gave AUROC 0.74 on 9,353 held-out requests. Prior to ML model development we built a simulation of the blood bank operation that incorporated returns to understand the scale of benefits of such a model. Using our trained model in the simulation gave an estimated reduction in wastage of 14%. Our partner hospital is considering adopting our approach, which would be particularly beneficial for hospitals with higher return rates and where units have a shorter remaining useful life on arrival.