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

TL;DR

A novel, machine learning (ML)-guided issuing policy is proposed to increase the likelihood of returned units being reissued before expiration to reduce platelet wastage 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.

Figures (8)

• Figure 1: The ROC curve of the trained predictive model on the test set, platelet requests for units required in 2017, and contour plots illustrating estimated wastage when issuing platelets using the trained predictive model within our proposed YUPR issuing policy. The receiver operating characteristic (ROC) curve on the test set plotted alone (a), and overlaid on a contour plot of wastage generated by our simulation of the hospital blood bank workflow assuming (b) the distribution of remaining useful life on arrival observed at UCLH and (c) the distribution of remaining useful life on arrival reported for a US hospital by Rajendran and Ravindran rajendran_platelet_2017. Under both settings, the predictive model could achieve lower wastage than an OUFO policy with no reduction in service level. A larger absolute reduction in wastage is possible in (c), in which wastage is higher due to the shorter average remaining useful life on arrival. Lighter colours indicate better performance. The region above and to the left of the contour for an OUFO issuing policy comprises combinations of sensitivity and specificity required for the YUPR issuing policy to incur lower wastage that OUFO. The ROC curve represents possible combinations of sensitivity and sensitivity that could be achieved by the selection of different thresholds for distinguishing positive and negative predictions, and the white star indicates the predictive performance that would be achieved on the test set when selecting a threshold to minimise wastage using the training set ROC curve.
• Figure 2: Summary plot of SHAP values for the 10 most important features in the trained prediction model. Predictions were made for each request, with patient-level features, such as "plt_count_value" (the patient's platelet count) based on the latest available data for the patient at the time the issuing decision is made. The feature importance values were computed using the training set. Each point represents an example from the training set. A large positive SHAP value means that the feature pushed the model output towards a positive prediction. A description of each feature is set out in Table \ref{['tab:feature_summary']} in Supplementary Note \ref{['app:ml_features']}.
• Figure 3: The order of events in one step of the simulated workflow, corresponding to one day. Emergency orders are only placed at stages 2 and 4 if there is insufficient stock to meet demand, and are made for one request at a time.
• Figure 4: Contour plots illustrating the performance of our proposed issuing policy with different levels of predictive model quality. The daily cost (a), service level (b) and wastage (c) assuming the distribution of remaining useful life on arrival observed at UCLH, and the corresponding metrics (d,e,f) when assuming the distribution of remaining useful life on arrival reported by Rajendran and Ravindran rajendran_platelet_2017. These plots show that under both settings our proposed approach can reduce wastage and cost, with no reduction in service level, relative to an OUFO issuing policy. Lighter colours indicate better performance. A perfect predictive model, with sensitivity and specificity both equal to 1.0, would be in the top left corner of a subplot. The region above and to the left of the contour for an OUFO issuing policy comprises combinations of sensitivity and specificity required for our issuing policy to perform better than OUFO. Each plot contains a labelled contour showing the performance for a baseline OUFO issuing policy and the colour map for each plot is centred on that contour. Subplots (a,b,c) are based on results from Experiment 2 and subplots (d,e,f) are based on results from Experiment 4.
• Figure 5: Impact of changing simulation input parameters on the the daily cost, service level and wastage when using an OUFO issuing policy and our YUPR issuing policy with a PPM. Values are the mean of each metric over the 10,000 evaluation rollouts, and error bars are the standard deviation of the metric over 10,000 evaluation rollouts.