The Future of AI in Cancer Care

Will Shapiro, vice president of Data Insights Engineering at Flatiron Health, shares insights into recent developments in the use of AI in cancer care and what advances are anticipated in this area. His presentation was part of the panel titled “AI in Cancer Care: Separating Fact from Fiction” at the 2023 Clinical Pathways Congress + Cancer Care Business Exchange.

Transcript:

Will Shapiro: My name is Will Shapiro and I'm the vice president of Data Insights Engineering at Flatiron Health.

What are some of the most important recent developments around AI and cancer care and what developments do you expect to see around the corner?

Will Shapiro: So at Flatiron, our mission is to learn from the experience of every person with cancer. And the way that we've done that traditionally is with human experts manually reviewing the charts of all of the patients in our population but there are limits to what that can actually accomplish.

And so what I'm really excited about and what motivates me every day is the ability to use machine learning and artificial intelligence to truly learn from all of the patients in our network because we have millions of patients in our network and that's something that humans cannot possibly manually review. So learning from the experience of every person in our network.

Can you talk a little about the importance of good ground truth and quality of data in AI models?

Will Shapiro: There is a common phrase in computer science, which is, garbage in, garbage out. And that is incredibly true in the context of artificial intelligence. If you don't have good ground truth data to validate against, how can you trust the output of a model? So when I mentioned that we have this team of clinical experts who go through and manually review charts, what that's actually given us over the past 10 years is an incredible set of ground truth labels so that we can have a high degree of confidence that when we're seeing that our model is correct, we know what it's being measured against.

I think one of the real challenges with things like ChatGPT in large language models is that they can hallucinate and they can produce very convincing answers that sound right, but are completely erroneous. So if you don't have a way to actually validate that the output of a model is correct and that you trust what correct means, there's no way to trust the model.

How can we ensure that the future of medicine is personalized for everyone, not just the targets of traditional clinical trials?

Will Shapiro: I think that the thing that's really key here is actually being able to learn from everyone. So again, what motivates me every day is Flatiron's mission, which is to learn from the experience of every person with cancer. And that means across socioeconomic bans and across race and ethnicity and genders and ages. All of these dimensions are critically important because if we are just learning from the traditional targets of clinical trials, that's what we'll continue to develop medicines for.

And as we know, medicines work very differently in different populations. So having both a representative data set to train on and ground truth to validate against is critically important to ensure that the models that you're building are not biased and ultimately that the drugs that are being developed are applicable for everyone.

What do you see as the future of AI in oncology specifically and in healthcare in general?

Will Shapiro: So there's a lot of different things that I'm excited about in this space, but I think that probably the thing that I'm most excited about is that AI has already done incredible things with being able to discover new molecules and new targets. But we're never going to be able to realize the promise of AI in drug discovery if we don't revisit how clinical development works.

And I think machine learning has a really important role to play here in making trials more dynamic and involving fewer patients and on more accelerated timelines. I think machine learning can play a huge role by learning from the real world and from patients as they are put on new therapies and dynamically updating the success or failure criteria of the trial.

You mentioned that the promise of AI to discover new molecules and targets is not going to be realized because of the way that the drugs are approved. Do you have any further thoughts on that? How can we achieve that?

Will Shapiro: I think currently the way the clinical development works is in phases and they are extremely time consuming and extremely expensive. And something that I'm really curious about is whether machine learning and AI can play a role in dynamically learning from the experience of patients in trials in updating the trial as it goes along. So rather than having a phased approach, there's a dynamic approach that incorporates the updated predictions of machine learning into the success or failure of the trial.