Algorithms are transforming health care. As medical devices are getting smarter, diagnosis is increasingly performed through machine learning, diseases are to an increasing extent monitored by algorithmic rules, and individual treatment options are more and more frequently decided by automated processes – with all the advantages and pitfalls that follow.
There are many advantages to this development. For one thing, automated decision making can free up scarce resources in the health care system. With aging populations and rising health care costs in many parts of the world, there is a huge potential for effective automation in terms of diagnosis, monitoring, and treatment. Furthermore, algorithms can potentially improve diagnostic and predictive accuracy, leading to fewer false positive and negative results, as well as overall better treatment outcomes for patients.
The increased use of algorithms in health care also comes with a number of potential pitfalls, however. The first is that errors in algorithms can be difficult to catch, especially in the case of “black box” algorithms, which can result from machine learning that can be difficult to asses and understand. Consider for instance this story about an algorithm developed to detect skin cancer through images. Although it was considered a huge success at first, it was later discovered that the algorithm used the presence of a ruler in images to determine cancer risk, as many of the images it was trained on contained a ruler, since dermatologists often use rulers to measure particularly concerning lesions.
The second pitfall of algorithms in health care is that automated procedures can lead to discrimination, systematically disadvantaging members of certain groups. Consider for instance this case about an algorithm developed to predict which patients would most likely benefit from specialized care, in order to avoid further complications etc. To find out which patients to designate to better treatment, the developers of the algorithm used past health care spending as a proxy for future needs. However, due to various reasons (having to do with historic and systemic racism), black individuals spend less on health care than white ones. Consequently, the algorithm was less likely to designate black patients to specialized care.
Another pitfall has to do with privacy, as health care algorithms often require access to large amounts of data that are extremely valuable. Especially in cases where collected and stored health data can be used for novel purposes over time, e.g. by running together mass scale digitized data collected from various different sources. Since algorithms designed for health care purposes often end up in the hands of private companies with much on stake, utilizing more and more algorithms in health care comes with a significant potential for misuse of sensitive information and breaches of privacy rights.
How to Mitigate the Problems
There are many possible ways to mitigate the potential negative consequences of using algorithms for health care purposes. Let me just name a few. Firstly, algorithms need to be developed on a foundation of scientific and moral principles. Developers need to engage with ethics to make sure that core values and principles are promoted and upheld through design and not as an afterthought. Moreover, regulatory authorities need to make sure that algorithms designed for health care are tested with the required scientific rigor.
Secondly, development of algorithms for health care purposes need regulatory oversight not just with regards to potential adverse events for individual patients but also with regards to collective adverse consequences. Ethics review boards therefore need specialized knowledge about the potential dangers of using automated procedures in the health care system. Manufacturers, on the other hand, need to collect real life data about their algorithmic devices, in accordance with risk, showing that their use do not have any adverse consequences in practice.
Thirdly, we need more literacy about algorithms in the general population. It is important for people to know about the ways algorithms are impacting health care (as well as crime prevention, banking, insurance etc.) so people can practice and protect their rights, as well as take part in discovering ways algorithms may have a negative impact on our lives. We need public education and outreach.
We do not need any of this because algorithms are inherently bad. On the contrary. Algorithms have the potential to transform health care for the better. To maximize the benefits and minimize the potential harm, however, we need to be informed about their effects in real life. We also need to monitor the current and future development closely. There are no algorithms (yet) that can do it for us.
Photo: Hush Naidoo, Unsplash.com