• 2 days ago
Patricia Brennan, Vice President, Science Technology, Chan Zuckerberg Initiative, Tilak Mandadi, Executive Vice President, Ventures and Chief Digital, Data, Analytics, and Technology Officer, CVS Health, Shiv Rao, CEO and Founder, Abridge Moderator: Verne Kopytoff, Senior Editor, Tech, Fortune, Co-chair, Fortune Brainstorm AI
Transcript
00:00Hello everyone. Thank you to our panelists for coming and talking to us today about a
00:04topic that's very important. Our lives depend on it. First off, can we perhaps start with
00:11you, Tala. CVS is a huge company doing a lot of things in AI. What is your favorite? Who's
00:19your favorite child?
00:22Favorite use case? Sure. Go ahead. All right. Look, I think anything we're doing with AI
00:32that makes our care and clinical professionals job easy is my favorite. So we have care managers
00:41that call people for next best actions and so on and so forth. We implemented AI based
00:49prep that saves them 40% of their time is now saved so they could spend more time with
00:57the patients. So those are the kind of things I like most.
01:01And then Patricia, to you. I can guess what your favorite thing is today.
01:05Yes.
01:06She has a bit of news that she has.
01:08Yes. Today I'm very excited that at CZI we're announcing and launching a platform to make
01:17kind of a one stop shop environment for biologists and ML developers to access models of biology
01:28across cell data, transcriptomic data, as well as imaging data. And these models are
01:34really designed to help biological research go faster and save time so we can get to understanding
01:42disease quicker and get to cures in a greater pace of time.
01:47Shib, could you tell us what a bridge does?
01:52Yeah, absolutely. So we're really focused on unburdening clinicians from clerical work so
01:59that we can bring the most important people in healthcare professionals and their patients
02:03closer together. So when you think about it, I'm a practicing cardiologist. When I see patients,
02:08after every single encounter, I'm always mindful that I'm really serving multiple stakeholders.
02:13I have to write a clinical note, for example, to give other folks on the care team, my colleagues,
02:19an idea of why I thought through this differential diagnosis, why I prescribed a certain diagnostic
02:24or therapeutic. I also have to think about revenue cycle because in this country, we're
02:28not compensated for the care that we delivered. We're compensated for the care that we documented
02:32that we delivered. So these documents have real implications on all things related
02:38to what keeps the lights on in the hospital. Then there's the third constituent, the third
02:42stakeholder who's the most important. There's the patient and the family member. If there's
02:47a term like transcatheter aortic valvuloplasty in my note, then I'm understandably going to get
02:52epic inbox messages or phone calls asking, what was that term? You never said it. I googled it.
02:56It sounds scary. So what we're focused on doing is taking from one conversation,
03:01from one conversation, all the information that we need to do a style transfer where we can serve
03:07all of those masters such that the clinician can just focus on the patient in front of them.
03:11Talak, your company is a partner of Abridge. How has it turned out? Have you encountered
03:18any wrinkles with the service? No. We work with Shiv and the team for the
03:28ambience AI in our clinics. It's working really well. It allows the doctor to pay more attention
03:38to the patient as opposed to turn around and start typing stuff, which is quite annoying.
03:42There's also AI that we use in our call centers that has the same, similar impact. So when you
03:56call somebody, when you call an insurance company, what have you, the most annoying thing that can
04:01happen is let me put you on hold while I pull your stuff together and 10 minutes will pass. That's
04:08just, that should not happen. So the way we are using AI in that context is that stuff would come
04:15together in real time. So the agent continues to speak. So really good, good, good positive impacts.
04:23Patricia, I mean, healthcare is a very regulated field when it comes to data and privacy. I mean,
04:29how much harder does it make to innovate in a field and in the field and make progress because
04:34of it? So I think, you know, with the research that we do and the products that we're focused
04:41on, it's really at the sort of very early stage and the basic research. So one of the things that
04:48we've really been able to do over the last several years is build up one of the largest integrated
04:55data sets of single cell data, which now represents over a hundred million cells. And those data,
05:03we've been able to, you know, make sure that you can see, you know, where those data have come from,
05:10what the studies, they link back to the original papers and whatnot. So it's really about making
05:15sure there's transparency to the data that underlies the models.
05:20And I hear that we have a new administration coming in and they have some ideas about AI
05:25and healthcare. What are you expecting, Tilap?
05:33Look, my view is pretty straightforward. I think there is an arc of how AI is going to transform
05:44healthcare. It's an inevitable arc. Administrations may change the timing a bit, may go
05:54here and there, some side branches. But the overall arc, in my opinion, is going to be in the
06:00direction of AI is going to augment clinical healthcare, never automate. AI is going to take a
06:12lot of administrative work off of the plate. AI is going to make human errors less and minimized.
06:22I think all those, and AI is going to help clinicians to get the information they need in
06:27real time. That's very difficult to get otherwise. So I think that's going to continue. That's going
06:33to be fine. Administrations are not. And Shiv, do you think that, you know,
06:38the Trump administration will be any better or worse than the previous one for startups?
06:44Well, I think the administration is absolutely going to promote responsible innovation. You think
06:49about it with healthcare in this country right now. Two out of five doctors don't want to be
06:53doctors in the next two to three years. 27% of nurses don't want to be nurses in the next 12
06:58months. So we have a public health emergency. We have to find a way to catch up to this emergency.
07:05And technology has to be a part of the solution. I think there's a recognition
07:08of that at every single level. Where I think it gets more complex is that not every use case
07:16can we sprint right into. And I think that there are the lower stakes use cases like clerical work,
07:21for example, where we're simply trying to take off people's plates the work that nobody wanted
07:26to do in the first place. And then I think there are higher stakes workflows or use cases around,
07:31for example, clinical decision support, where we necessarily need to have some friction,
07:35some good friction in those workflows in order to make sure that we can deploy those technologies
07:40responsibly. So I think at this moment in time, and what we've seen over the last two years,
07:44is that those solutions, those products, those technologies that are on the safer end of that
07:49spectrum, they've scaled. And I think that they'll continue to scale probably even more quickly over
07:54these coming years. And at the same time, I think we also as an industry have to start thinking
07:59about what those next innings look like, even around clinical decision support,
08:02and how we can responsibly set the scaffolding for us to go after them.
08:06And to look, insurance claim denials have reached the headlines following the murder of the United
08:13Healthcare CEO. And I should say, CVS owns an insurance company. Does AI play a role in the
08:25insurance claim process? And is it possible that AI could help with the error rate of denials?
08:34First of all, it was a horrific thing that happened and sincere condolences to the family
08:39and United Healthcare. Look, what do people want from an insurance company? They want transparency.
08:51Why did that decision happen? And an explanation that is in real time, not later. Transparency to
09:02benefits, what is the coverage? And transparency to cost, what is my out-of-pocket cost?
09:10And an easier way to present a case. These are fundamentals that people want in an experience,
09:23healthcare experience. And insurance companies work very hard to provide that. AI is going to
09:31help immensely in all those things. Speed up, you know, why can't we have real-time
09:39pre-authorization at the time of care? The doctor is sitting there with the patient.
09:45They should be able to do a pre-authorization at that time of care, as an example, right?
09:53Exposing the insurance contracts, especially in the commercial space, tend to be pretty complicated.
10:00So AI can help navigate that very quickly in real-time to show what is covered, what's not
10:07covered, and so on and so forth. So yes, I think it's an ongoing process. We've been working on
10:14this for a while. And AI is going to help even more to address those things, the transparency,
10:20the explainability, the ability to have a presentment of case and so on and so forth,
10:27make it easier. In a few minutes, we'll go to audience questions. So please
10:32come up with a great one for us, please. Patricia, you know, everyone knows that health
10:38costs are soaring. Do you think that AI is a solution to at least some of that?
10:45So I think that AI can help accelerate how basic research is done, for example,
10:55or save time across the entire sort of life cycle of sort of a bench-to-bed, as it were,
11:01or a bench-to-care application. And so I think there are opportunities to save the time of a
11:10researcher by bringing more computational capabilities, more tools to bear on the work
11:17that they're doing and get the research out faster, get it to the clinician or the application
11:23healthcare faster. There are some, you know, costs associated with scaling up and, you know,
11:31actually applying AI in the research process, but I think ultimately it's about saving time.
11:40Maybe we could go to questions here. We have somebody
11:44with a paddle. Yes, sir, go ahead and state your name and affiliation and go ahead.
11:50Venk Vardhan, co-founder and CEO of Nano, where we're a home-based digital diagnostics platform.
11:56Shiv, you touched upon this point that I thought was interesting. I think we can all agree in
12:00healthcare that AI will be great for workflow optimization, logistics, things like that. That's
12:05not as much as the next bar of diagnostics and therapeutics. What is the panel's opinion on
12:11some of the new aspects around that latter bar around validation liability, this concept of
12:17assurance labs versus FDA, reduction of false positives and false negatives? Obviously, if we
12:22can corner and figure that part out, there's a real unlocking revolution in healthcare. So,
12:28curious on you guys' thoughts there. I can kick things off, maybe to Locke's point as well.
12:34The only currency that matters in healthcare is trust. The stakes can be so high, especially
12:39in care delivery, and if trust is some combination of transparency, reliability, and credibility,
12:46I think what a lot of us are trying to get our arms around is what does transparency but
12:51also reliability look like? If you're using a solution in the wild, how accurate is it?
12:58How reliable is it? And how do you measure that? So, in our space, we have technologies that
13:07understand language and try to recognize the words that people are saying, regardless of what
13:10language they're speaking in, even. We have technologies that try to sort of abridge the
13:14conversation, sort of distill the essence and structure that data, as well. We have technologies
13:19that try to summarize, as well, for all the different audiences. And so, for some of these
13:24challenges, it's pretty straightforward for us on our machine learning team, on our science team,
13:29to sort of understand, okay, that's an F1 score. We're going to be measuring precision and recall,
13:35or okay, for this, we're going to be measuring medical term recall and word error rate, for
13:41example, for speech recognition. But for some of these technologies, it's a little bit more
13:45challenging. For example, on summarization, how do we measure ourselves? How do we benchmark
13:50ourselves? And so, one pursuit that we've invested a lot of time and cycles into is just being able
13:59to publish benchmarks around where we are today, being able to also give people a sense of how
14:06we assess quality ourselves within the company, and also trying to expose these playbooks out to
14:12the broader industry so that we can all learn from each other. When I think about these assurance
14:17labs, I think, to some extent, I see them sort of pulling inspiration from what a lot of different
14:22parties are publishing. And will we come up with one standardized sort of checklist on how to think
14:28through any given technology and any given use case? I think it's going to be a challenging
14:34task. And I think one of the reasons why is that the ground is moving underneath us so quickly.
14:40These technologies are evolving so quickly that it's hard to sort of take a snapshot of any
14:46given solution, much less problem, and sort of call it done. We take a lot of pride that
14:52Bridge on Mondays should be great, but it should be even better on Tuesday, Wednesday, Thursday,
14:57Friday. We also take a lot of pride in sort of moving up the stack and expanding the scope of
15:02what we're building into other use cases. And so how do we create these standardized
15:08benchmarks for all these different technologies and all these different workflows? It's tough,
15:13but I think we need some kind of high-level rubric probably across the industry to just
15:17sort of guide us. My sense is that it's going to come down to proxies for trust.
15:23Anyone else with a question? Go ahead. Do we have a paddle up here?
15:32Hi. Lorraine Fraser at Columbia University, and I happen to be the dean of nursing. And I
15:42agree and disagree with you. I think the only currency is also the dollar and trust.
15:49For hospitals, for insurance companies, let's be honest. How do you bring ethics
15:56into developing these models of AI? I would like to understand that.
16:02Start with you, Patricia.
16:04Sure. I'll start with that. I mean, I think that ethics and all aspects of trust,
16:11you mentioned privacy already, having transparency into the data. Actually,
16:17considering even before the research into the model starts, what are some of the
16:23scenarios and use cases that are really trying to be enabled by them and bringing those to bear
16:29throughout the process? Benchmarks are a huge part of this. There's, I think, a real opportunity and
16:34gap in the biomedical research field as it relates to sharing and bringing some consistency to how
16:43benchmarks are made available across the research lifecycle. So I think it really is about applying
16:52some of these core principles of privacy, data transparency. Openness is another key aspect to
17:00this. If you could probably take that offline. We're out of time here. I want to thank our
17:08panelists for a fabulous conversation. Thank you.

Recommended