“Enigma is an extremely well-designed machine. Our problem is that we are only using men to try to beat it. What if only a machine can defeat another machine? ” Alan Turing, in an exchange with Commander Denniston, at Bletchley Park
Ethics is increasingly important as a domain in artificial intelligence in medicine. This area is especially relevant now with COVID-19, with a myriad of issues ranging from privacy and confidentiality in digital contact and tracking data to real-time decisions in the ICU setting regarding patients in the midst of scarce resources.
One framework that we can use to discuss ethics in artificial intelligence in medicine as it is related to health disparities is to deconstruct the machine learning workflow into its three main constituent parts:
If the biomedical data has embedded inequalities, then the algorithms will automate and even perpetuate these inequalities. Biomedical data curation is arguably the most difficult steps to get right in order to minimize bias, so it is essential to invest time and resources to ensure that the biomedical data is without inherent and/or unintentional bias. As a tribute to the aforementioned quote, there is ongoing work on using transfer learning for neutralizing biomedical data inequality.
The algorithm also needs to pay special attention to any possible bias based on sex and gender as well as ethnicity as many models lack sufficient diversity and therefore cannot generalize to other health systems. To minimize healthcare disparities, unbiased AI models are essential prior to deployment, and perhaps there is a future role for a specially-designed algorithm to check for bias in other algorithms.
While avoiding bias in biomedical data is perhaps the most challenging part, real world deployment is maybe the least predictable aspect of the entire workflow in avoiding bias. How anyone chooses to use the model and to monitor the output is without much standardization or regulation at present, so human oversight is very necessary to minimize the presence of bias amongst these models.
In short, we learn much about ourselves in this process of scrutinizing artificial intelligence for bias. Therefore, the three takeaways in ethics of artificial intelligence:
- Start with biomedical data without data inequalities;
- Consider AI methodologies to help reduce bias in the algorithm; and
- Implement human oversight as a continual monitor for bias.