AI/ML Bias
Regulatory and Ethics
Date: December 8, 2021
Time: 4h30PM EST
Meeting Chair(s): Yann Joly
Attendees [Name(Affiliation)]:
- Lindsay Smith
- Kristina Kekesi-Lafrance
- Anna Lewis
- Yann Joly
- Maili Raven-Adams
- Supriya Kapur
- Tina Hernandez-Boussard
- Mollie Minear
- Ma’n Zawati
- Judit Kumuthini
Passcode: G21$3TdX
Agenda Item | Speaker | Time | |
1.0 | Introductions | Yann | 10 min |
2.0 | Clarification on topic and terminology | Tina | 15 min |
3.0 | Potential outcome | Group discussion | 15 min |
4.0 | Choice of Co-Chair | Yann | 10 min |
5.0 | Next steps | Yann | 15 min |
6.0 | A.O.B. | Group discussion | 5 min |
Zoom Information:
Join Zoom Meeting
https://us02web.zoom.us/j/88019377709?pwd=M0RLY2xtN2NuSmhjTVV3WmJzSDFNZz09
Meeting ID: 880 1937 7709
Passcode: 611935
Find your local number: https://us02web.zoom.us/u/kcLeus0yO
Introductory meeting: Oct. 14, 2021 @ GA4GH Connect
AI/ML Bias
This session provided the current landscape of AI/ML bias, beginning with examples of AI/ML bias, biases that can occur throughout the AI lifecycle, types of biases in the data, model evaluations, and the importance of building trust through transparency. Afterwards, the group conducted a discussion on opportunities for AI/ML bias standards within GA4GH. A new cross Work Stream initiative will be created to continue discussions and work in this space, and the team will aim to conduct a survey of AI/ML use across GA4GH.
Minutes:
1.0 [Introductions]
2.0 Clarification on topic and terminology
Goal: the need to seek a case study that we could use to test our ideas. We want to provide practical and ethical tools/guidelines. Then we can zoom out.
Who is actually using current models (and its generated data) and in which context?
What should be our starting point to set some standards or guidelines at every point of the life cycle of the data? What are the models used by the group and where were they built, public/private sector?
Where is this group best equipped to contribute? Data creation/acquisition stage: we often don’t know what the data will be used for downstream.
Maili: I guess, as AI is continuously emerging it would also be important to ask where they groups may intend to use AI in the foreseeable future - as they may not have adopted it yet
What is the definition of AI we are using? Using health data to train a model for evaluation.
We are learning from an existing set of data, which is also true for polygenic risk scores, and that is not AI. The PHG foundation may have some starting points for definition.
Important to survey the community to understand where they are getting their data from and what they wish to use it for. It will also tell us where there are opportunities to best use this data.
Problems that practitioners are already aware of: downstream statistical interpretation effect, 1000 Genomes project where they classify as Indian population people that are actually Sri lankans, even if they share very genetic makeup, still huge differences between both populations and should not be considered the same. Nonetheless, people still use 1000 Genomes. Africans: most of the tribes have their own identity (masks are like GPS from where you come from/your ID, has 50% of the phenotypic information we collect).
How distinct is the question of AI and bias vs diversity? Overlapping efforts? The majority of the bias stems from the lack of diversity of the data we are collecting, but that is only one component. If we can build on what others are doing, that would give us an opportunity to lead in other areas/leverage what was done in that space and build on it.
Diversity and Datasets group: danger is dividing our forces. Most of the people in the group here are also interested in other REWS’ initiatives relating to diversity. Either we combine certain efforts or make sure we are distinct. No conclusion for now.
One way to distinguish both groups would be to identify outcomes so we can either co-develop them or combine them.
Should see first what comes out from the survey. Maybe some elements will stand out as more specific to this group.
Can Tina or Supriya attend the next REWS meeting to make sure efforts are not overlapping and triggering the interest of others in our initiative. Then we can see if there is an interest for us to stay in a separate working group. If we merge it’s only because there is either value to it or we avoid duplicate work.
Another example without the lack of diversity: equitable access issues? Is there bias in how we can make sure an algorithm is transparent to researchers? For example making sure it is explainable to people without an AI background?
Maybe the introduction of diversity in the data at the data collection point should be undertaken by the diversity group. If we look into bias at every stage of the data, we may get lost.
Maybe more focused/narrowed to say how not to use data instead of how to use it.
The other piece we have not mentioned yet is that there are a lot of guidelines out there to measure bias and to compare what is the bias and bias from what. How do I know if data is biased and fit for purpose?
3.0 Potential outcome
Red flag publication? Otherwise, it can be extremely broad.
Will depend on the survey and the discussion with other REWS’ working groups.
2-3 pages which summarize the data bias issues that come up in AI that are obvious/not so obvious and that can be our way to dive into that work, from the diversity group (as a starting point) and distinguish ourselves.
4.0 Choice of Co-Chair: please contact rews-coordinator@ga4gh.org
5.0 Next Steps
Survey: Anna will draft 4 questions as a starting point to ID AI in our GA4GH community.
Anything in the public domain + willing to have a conversation with us ?
Would also be interesting to know what is the purpose of using AI: academic publications, or other use cases for research.
Are people using/creating their own tools or using top of the shelf tools? Have they gone through any regulatory process at all?
Tools that are deployed at all stages of the pipelines or directly to the patients in the clinical setting? Think of: researchers, insurance companies…
Connect Tina and Supriya with the rest of the GA4GH community, REWS and other working groups within such as diversity. Will help us untangle what everyone is doing.
Identify a definition for AI: start from the PHG foundation’s definition.
Next meeting: monthly basis + Doodle poll in the New Year.
6.0 AOB
RESSOURCES
https://www.phgfoundation.org/report/artificial-intelligence-for-genomic-medicine
This was a review of where AI is being used and is headed in AI/Genomics
Roxana Daneshjou, MD, PhD1,2; Mary P. Smith, MD3; Mary D. Sun, MSCR4; et al Veronica Rotemberg, MD, PhD5; James Zou, PhD6,7,8
JAMA Dermatol. 2021;157(11):1362-1369. doi:10.1001/jamadermatol.2021.3129