Kurt J.G. Schmailzl, ccc. Center for Connected Health Care UG; Carl-Thiem-Klinikum (CTK) gGmbH, Cottbus
Algorithms have long been used in medicine to support diagnostic processes and guide therapies; simple examples include analysis programs for ECGs based on the measurement of time intervals and voltage amplitudes."Risks" existed only in incorrect measurement, and the prospects were clear: manual measuring could be dispensed with. When we use algorithms to assist diagnostic processes and to guide therapies, the picture changes: When we train appropriate algorithms, doctors and technicians work closely together: The technician proposes an algorithm, which is tested on a data set, and the physician assesses to what extent the algorithm meets the same criteria as a human would have preferred. Even if the medical professional relies on guidelines, his view of the case is not necessarily the view of 2 or 100 or 1 million medical professionals on this particular issue. However, if his view becomes the curriculum of a new algorithm, then that single view prevails. This single view could, however, be distorted by consciously or unconsciously considering or not considering gender or social aspects. For example, some diseases affect women and men differently. Perhaps they simply affect older and younger people in different ways. Or members of higher or lower social classes. Or members of different professional groups. It is likely that the genetic background of the person concerned plays a major role. And all of these factors affect the diagnostic accuracy and appropriateness of a therapy recommendation. One could try to make the algorithm more complex so that as many of these co-factors as possible are taken into account. The training of a proposed and negotiated algorithm is the critical moment where bias is introduced. As soon as we finally allow algorithms to decide on our further actions, we have left behind the moment when we could still make corrections. Action-guiding algorithms in medicine will no doubt increasingly decide on health and illness in the future. They will possibly do so without any human being having to be involved in this decision-making chain at all. And we will find it a relief in our everyday life. In other words: we need repair mechanisms that allow a change of perspective: just as we are asked to change a password at certain intervals.