5.4 Discussion

Risk of iron-related adverse outcomes at follow-up donations can be estimated as a function of the return time using data available at an index donation. Estimated risk decreased precipitously for most donors if they waited longer to return, suggesting that tailoring donors’ IDIs to individual donors’ risk profiles may be an effective strategy for managing risk of iron-related adverse donation outcomes without unduly restricting return donations from low-risk donors. Risk for some donors remained high even 250 days later, suggesting that these methods may also be useful for identifying donors who may be poor candidates for repeat blood donation. Including ferritin as a predictor improved risk estimation, particularly with respect to estimating risk of absent iron donations.

Our analysis has several limitations. The RISE study asked participants to commit to frequent blood donation and targeted recruitment to achieve proportional representation based on gender and donation history. Further study is needed to assess the generalizability of our prediction model’s performance to a general blood donor population. Additionally, the outcomes we estimated at the population level are specific to the RISE cohort. In particular, the baseline rate of adverse outcomes and the reduction in supply introduced by tailored IDIs may be lower in populations with lower return rates. Another limitation of the RISE data is that ferritin and other biomarkers were not measured for all follow-up visits. We factored this into our analysis by assuming these biomarkers were missing at random, but this may not be the case.

Two other limitations must be overcome before tailored IDIs can be implemented in practice. First, a decision-maker must identify risk thresholds for each adverse outcome in our method. Experts may not agree on the level of risk that is acceptable and how the sufficiency of the blood supply should be weighed against risks to donors. Further work is needed to understand these trade-offs and identify reasonable risk thresholds. Second, this method may face significant barriers to implementation. Sophisticated machine learning techniques for decision-making require technical expertise to develop and maintain and are opaque in the sense that humans cannot readily understand how the system arrived at a decision. In a growing literature on interpretable machine learning, methods have been developed for constructing simpler decision rules that sometimes perform on par with advanced machine learning techniques [146–150]. Further work is needed to assess barriers to the adoption of the tailored IDI method developed here and to determine whether simpler decision rules might be easier to implement and perform similarly. Despite these limitations, our analysis suggests that individual risk prediction could be a useful tool for ensuring a sufficient blood supply while managing iron-related risks to repeat blood donors.