Case Study: Artificial Intelligence Bias in Medicine
Lisa Falco is the Director of Data Science at Ava AG, a San Francisco and Zurich based firm dedicated to using artificial intelligence (AI) to offer women insights about their fertility and reproductive health. Lisa will be speaking on March 28, 2019, Techdebates.org in Zurich discussing whether or not AI can be trusted in the face of biased data. I had the chance to briefly talk with Lisa to get a sneak preview of her thoughts on the topic and some of her views on AI progress in general.
AI is in its early days. What have these early days of AI looked like for Ava?
Lisa: Our solution helps women who are trying to get pregnant track their monthly cycle to predict the days they will be the most fertile. While our users are sleeping, the Ava bracelet’s sensors collect data on key physiological parameters. Ava’s algorithm then detects their fertile window, physiological stress level, sleep quality, and other health factors. Other methods of fertility tracking rely on only a single parameter, and thus, they cannot generate an accurate and complete picture of fertility conditions in real time. Ava, in contrast, considers many parameters and provides biofeedback based on each woman’s unique current conditions.
Some companies promote different versions of easy plug-and-play AI. But our experience indicates a lot of preparation and customization is still necessary—especially on the data side. I think Ava, and AI-driven companies in general, will become increasingly adept at using more standardized data constructs, but we’re not there yet.
For our solution to work effectively, we train our AI using not only the data we are collecting from our users but also evolving medical knowledge. We work with doctors and spend a lot of time reading medical literature to constantly improve the information we’re feeding our algorithms.
When we first started the company, we didn’t have much data to work with, so our technology was not really AI-driven. Rather, our first algorithms were expert algorithms, meaning that they were designed by humans based on medical research. As our user base grew, we had big data to work with—but even with big data, medical knowledge still does and must substantially influence the algorithms.
As we collect more and more data from users, our algorithms are constantly being improved by the AI; freeing us from reliance on a human-derived model. By using AI, we can capture much more variety relating to women’s complex physiology. Now, we have this big data pool with many, many data points. The human mind simply can’t capture all these variables and intersect them at once. This is the power and the value of AI.
What ethical considerations are we dealing with here?
Lisa: Ava manages important ethical concerns as well as regulatory considerations. We already do a great deal of validation work for Health Authorities, such as the Food and Drug Administration (FDA) and the regulatory constructs we are operating under will continue to become more demanding. First, it’s very important to develop your algorithms properly and to do the correct splits between the training sets and the validation sets. Next, we do an external validation with our clinical team. From a clinical perspective, we’re testing on data that we have never seen before. In the end, it’s vital that our algorithms account for the wide variety of women we hope to include in our user base.
Being in the medical space is an interesting journey for a technology company. As a tech start-up, the initial focus is on being fast—the faster the better. But as the technology moves into clinical application, the pace slows way down and there is a lot of focus on validation and regulatory approval. Although it can be challenging for a tech company to accept this slower pace, it is necessary because we are dealing with patient health.
What do you think is possible for AI in the short term, and in the longer term?
Lisa: I think most well-defined, fact-based decisions and tasks will eventually be solved by AI. But primarily, it will serve as an additional input to human decision-making. Particularly within the medical field, I don’t think anyone is striving to replace doctors. Rather, we want to assist them in making better and faster decisions.
An interesting value AI-driven medicine could deliver relates to healthy test results. We all go to the doctor for preventive health checkups: mammograms, yearly physicals, things of that nature. A very large portion of these tests reflect a healthy patient—no action needed. These cases could be easily sorted out by AI, allowing doctors to focus on solving problems and helping patients in cases where it really matters.
What do you see as the biggest risks for creating biases in AI data and AI algorithms?
Lisa: AI learns from the data it is trained with. Not necessarily in the medical space, but in general, the data used by many industries are biased toward the way the world is today. To be more specific, most industries are led by white men. If we are not careful, we will train AI solutions with input data that might reflect the biases among today’s industrial leadership. This may not be reflective of the actual population or a more egalitarian construct.
You also must ask whether your training data and algorithms are fit for purpose. For example, our current user base is women trying to get pregnant. However, we are also developing algorithms for contraception purposes. We have had to ask ourselves whether data we’ve collected from women trying to get pregnant is truly representative of women who want to avoid pregnancy.
Perhaps the women trying to get pregnant are of different ages; they could have different lifestyles; maybe one group has better overall health than the other or one group has greater ethnic diversity. You simply cannot assume that these two groups are the same because they likely are not. But without recognizing and compensating for these differences, you can inadvertently create a great deal of problematic bias.
As AI advances, the ethical considerations of potentially biased data will become more entrenched and will likely become increasingly difficult to fix. How might we best manage this issue moving forward?
Lisa: We need to increase the diversity of people working within the AI field. I once visited Google’s headquarters, and as a male Google team member and I walked through the campus, I noticed that there were only guys—at least the vast majority of them were. A bit into our walk and conversation, he chuckled and commented that Ava would never have been invented at Google. No one would really care much about the issue of fertility. More specifically, because Google is so male dominate, no one would have thought about AI’s ability to solve the problem.
There’s a great book by Hans Rosling titled Factfulness: Ten Reasons We’re Wrong about the World—and Why Things Are Better Than You Think. In the book, he discusses how we make biased conclusions about how things work based on our specific experiences. Rosling shares a story about a visit to Africa with some of his students. He and a few members of their group were standing in an elevator; however, a student from Sweden was running late and saw that the elevator doors were closing. As she ran toward the elevator, she put one of her legs between the doors; as in Sweden, and in the Western world in general, elevator doors pop back open when faced with an obstruction. But the elevator doors in this part of Africa didn’t have that safety mechanism, and she almost lost her foot.
The point is that something may work one way in one part of the world or among one group of people, but these experience-based conclusions cannot be universally applied. We must be incredibly careful to avoid biases in AI data and within the algorithms themselves. Increased diversity among AI professionals is one important factor.
To learn more about rapidly evolving AI technology and the ethical considerations that accompany these advancements, join us for the March 28, 2019 TechDebate in Zurich, Switzerland. Lisa and Nicolas Perony, CTO, OTO.ai, will continue this fascinating discussion.
My company, Sphere Software, is a sponsor and organizer of TechDebates.