Artificial intelligence use in clinical decision-making:allocating ethical and legal responsibility

[purplecat]

The workshop last week included a poster session and, as is the way of these things, one can't always get to speak to all the poster presenters one would like to. This one had included handy references and QR codes so I just snapped the poster and moved on - now need to decide which papers to add to the large virtual "to read" pile.

Comments

[vivdunstan]

I've only encountered posters once at an academic conference, the SHARP book history conference at the Bibliothèque Nationale de France. Posters aren't generally a thing at humanities conferences, so this was a new experience for me! Though more familiar to Martin who was with me that day. I had to drop out of my CS PhD too soon before running into them at science conferences. We did manage to speak to some of the people at the Paris poster session, but spent much of our time photographing pictures, and especially using Google Translate app to translate the ones in French!

[purplecat]

Yes, we had a whole "explaining posters to philosophers" pre-amble in the workshop which, to be honest, I felt rather over-sold posters as a communication mechanism!

[a_cubed]

On a more practical angle for related research you might find the work of Dr Sachil Singh at York University (the Canadian one) of interest. https://health.yorku.ca/health-profiles/?mid=2080700 . He's been looking at embedded racial (and other) bias in existing CPD systems which use "AI" to generate training material based on academic publications. There are multiple problems he's found in practice. Many medics are using these training systems as close to decision support systems (and the companies that produce them are obviously interested in getting into that area properly). The research projects may be biased (lack of diversity in test patient data for example) or even the literature on a particular area may be biased. Finally, their "AI" interpretation of the data seems to introduce new biases by misinterpreting statements or numbers in the literature about racial make-up of patient base, such as interpreting a study trying to address systemic lack of study of non-white patients as meaning that non-white patients have a greater prevalance of certain conditions.

He gave a talk here a few months back, though I'd already met him at a workshop when he was a PhD student in surveillance studies.

[purplecat]

What always surprises me (though it shouldn't) is how unaware so many people are of algorithmic bias. I feel like the AI and Ethics crowd have talked about almost nothing else since Buolamwini's Gender Shades paper what, five years ago now, and yet people are still unthinkingly deploying these systems.