Fair by chance? On the use of algorithms in therapeutic decisions

Predictive tools made possible by advances in machine learning techniques may help clinicians make more accurate decisions about who should be allocated costly therapies, such as immunotherapy, which only work on a relatively low proportion of patients. In this article, I argue that a fair decision procedure must recognise each patients' chance of responding well. To do so, the procedure should not apply a fixed threshold to probability scores. Rather, each patient should be given a chance of being allocated the therapy matching her probability score. An important consequence of this conclusion is that the fair use of algorithmic scores may not guarantee that the therapy is allocated to those who will respond well to the highest possible degree.