Even as artificial intelligence (AI) has become an increasing focus in many areas within U.S. healthcare, clinicians, informaticists, and others are all discovering just how difficult and time-consuming the development of AI algorithms is actually turning out to be. By one estimate, it can take an investment of three years and $5 million per model to achieve generalizability. One of the biggest roadblocks to AI’s advancement is the quality and quantity of data needed to properly train algorithms, as an algorithm that performs well on local clinical data often fails when exposed to diverse real-world data. Acquiring the volume of diverse, real-world data needed to retrain algorithms is not only complex, but time-consuming and expensive.
But at UCSF Health, the academic medical center organization based in San Francisco, clinician and informaticist leaders have been moving ahead to log real gains in this key area. Indeed, the work of the UCSF Health leaders has led to a spinoff company called BeeKeeperAI™, which was created at UCSF’s Center for Digital health Innovation (CDHI) to solve the healthcare data access problem.
As BeeKeeperAI’s leaders note on their website, “As we developed clinical AI at CDHI, we were experiencing the same data acquisition challenges that currently plague all healthcare AI algorithm developers and researchers. While our algorithms were developed on high quality data sets and continued to perform very well on local clinical data, they did not perform as well when exposed to diverse real world data. Imaging-based algorithms, in particular, suffered when they were tested on images acquired on different and older equipment and techniques. Acquiring the volume of diverse data to retrain the algorithms to achieve the desired performance was complex and very time consuming. It quickly became apparent that both academic and industry AI developers were struggling with exactly the same issues and that a new solution to the data access challenge was needed if AI was going to live up to its promise to improve healthcare.” BeeKeeper will be fully spun out during the first half of this year.
One of the key leaders in UCSF Health’s work in artificial intelligence has been Aaron Neinstein, M.D. Dr. Neinstein is an associate professor in the UCSF Division of Endocrinology, vice president of digital health for UCSF Health, and senior director at the UCSF Center for Digital Health Innovation, which he helped to co-create in 2012, and which currently involves a multidisciplinary team of 50-plus designers, project managers, data scientists, and engineers.
His digital health resume spans leading UCSF’s ambulatory Epic EHR (electronic health record) implementation, development and deployment of numerous digital health solutions transforming digital patient experience and virtual care delivery across UCSF, research and advocacy advancing US federal health policy in interoperability and patient data access, and co-founding of Tidepool, a non-profit creating open-source software to empower people with diabetes. At the UCSF CDHI, his multi-disciplinary team focuses on advancing patient experience and digital transformation of care delivery, including the use of telehealth and remote monitoring technologies to advance more connected care. Dr. Neinstein was an inaugural inductee as a Fellow of the American Medical Informatics Association (FAMIA). Board-certified in endocrinology, clinical informatics, and internal medicine, he maintains an active clinical practice focused on the care of people with diabetes.
Dr. Neinstein spoke in December with Healthcare Innovation Editor-in-Chief Mark Hagland about the AI initiative at UCSF Health. Below are excerpts from that interview.
How UCSF is approaching AI development overall?
We are enthusiastic about AI in the healthcare space. We in partnership with GE Healthcare a few years ago built the Critical Care Suite—a series of AI algorithms that sit on portable x-ray machines and facilitate quicker clinical care and faster triage of important clinical findings on chest x-rays in the course of care. For example, if the breathing tube is in the wrong place, the x-ray machine can inform the team, or if there’s a pneumothorax, which is a collapsed lung or about to be collapsed, which can be a life-threatening emergency, the x-ray machine identifies it immediately as the x-ray is being shot, so rather than waiting for the radiologist, it alerts the team immediately.
This program both alerts on the machine and then orders the studies with urgent findings, it pushes those studies to the top of the list. Our team developed this from 2016-2019. It was licensed by UCSF to GE, and the first x-ray machines went out into the field in 2020. I believe this is one of the first on-device AI algorithms. So that was one early foray into a clinical use of AI. In general, our working hypothesis is that while we’re very excited by the direct clinical care use cases for AI, we also think there are huge opportunities for AI to drive operational and business efficiency. People get so excited by the AI doctor that it’s a bit underappreciated about AI healthcare.
Meanwhile, we’ve also been working on a few things with Philips. One is that we get 1.5 million faxes a year at UCSF, and faxes are still extremely highly used in healthcare institutions. And while we’re trying to continually move toward more electronic communication, people continue to send us faxes, and we can’t stop that. Meanwhile, about one-quarter of those faxes involve referrals—about 200,000 referrals that get faxed to us per year. And each one of those referrals involves diagnoses for such conditions as cancer, rheumatoid arthritis, or type 1 diabetes. And in the fax-based world, there are a lot of delays involved. So we’ve created a referrals-based automation so that the AI algorithms are reading the faxes and can pull the information into the EHR so that our referrals are far faster. And it allows our staff to spend a lot less time typing things from faxes. Huge operational efficiency and patient experience benefit from leveraging AI in this sort of basic and boring way.
When did this go live?
The first version of the software without AI went live two years ago; the AI is going live over the next few weeks. H2O.AI and Philips and we, all three collaborated on this, and are working on different parts of the software.
What is the mechanism involved?
That is where things get fascinating. In the vast majority of cases, faxes are paperless, so senders are sending an efax. So it’s essentially a plain-text PDF being sent to us. And so it’s a paperless interaction, but because it creates essentially an image file for the exchange, the data are not digital, so when we receive the image file, the AI is reading the image and translating the data back into digital format that can be translated into a referral. We all think of faxes as paper; in fact, these days, they’re largely digital—indeed, the overwhelming majority are efaxes—but they still need to be translated.
What gains do you expect from this?
Significant reductions in referral turnaround time, meaning, if your primary care doctor sends us a referral, we’ll process that referral and get you into the scheduling pipeline almost immediately. I think similar to the x-ray machine advance, computers still process information sequentially, but they do it so quickly that, on a human level, it’s as though everything is being processed simultaneously. And if we get 500 referrals in a day, it’s possible that referral number 499 is the most important referral. Well, AI can start to learn patterns about which of the referrals have the highest degree of urgency, and can move those referrals to the top of the pile.
Training AI algorithms to recognize abnormal patterns is happening the fastest, right?
Yes. If there’s a pile of 100 tasks, whether you’re a radiologist processing x-rays or a staff person processing referrals, every human will work tasks in some random or trained order. But if AI can resort those, it’s not trying to replace a human or get a diagnosis right where the stakes might be a lot higher. By triaging information, it’s moving urgent things to the top of the pile.
We’re involved in two other program areas to tell you about related to AI. One is another one we’re working on with Philips, per its Patient Flow Capacity Suite software. We think of it as “care traffic control”: understanding where a patient belongs in the hospital system—in the ED, on a regular hospital floor, in the ICU, in a step-down or transitional care unit? Is it time to be discharged to a SNF? Which facility has the beds? We’re starting to test, with a series of algorithms, where a patient should be. So we’ll be managing bed availability and staffing. It’s really a big deal with COVID now. If one hospital has better nursing coverage, maybe a patient should go to a different facility. We’re hoping to deploy that in the next four to six months.
Bed capacity management and unity capacity management are becoming absolutely crucial now, correct?
Yes, capacity management really is a huge deal. Most hospitals are developing some kind of command center. We think there’s a lot of opportunity for big data and predictive algorithms to do a way better job than we can. You can imagine ingesting data about nursing staffing, capacity issues, and you can imagine how that will help, in terms of nurses.
And, in one other area, we’re spinning out BeeKeeperAI from UCSF. Beekeeper is tackling the problem of the fact that an algorithm is validated on one institution’s data, and then, when you try to move it to another institution, it doesn’t work, because the algorithm hasn’t been trained on a broad enough data set and the data are different in different organizations; so pooling together lots of healthcare data from lots of different institutions is a very risky endeavor. So BeekeeperAI does zero-trust data management. Let’s say, for example, that NYU has an AI algorithm we might want to test. They’ve validated it on NYU data, but before they could go for FDA approval or commercialize it, they’d have to test it on many other institutions’ data. So rather than creating a situation where they’re giving all their data to another institution or vice-versa, BeekeeperAI brings the data into a secure enclave, brings in their algorithm, tests their algorithm against the other institution’s data inside a black box, produces a validation report, and neither side sees the other’s data, and they go about their business. NYU never sees UCSF’s data and UCSF never sees NYU’s algorithm; it allows for far faster testing on algorithms on a broader set of data.
When was BeeKeeper founded?
BeeKeeper was created through the course of 2020 and 2021, in partnership with several companies—Microsoft, Intel, Fortanex, etc.—and it’s currently getting venture funding and will be spun out of UCSF in early 2022.
What have the biggest lessons been learned so far about AI, by your team, and more generally?
One of the themes you’ve seen across these use cases is that there’s a lot of administrative burden in healthcare. Doctors and nurses spend a lot of their time on administrative tasks rather than on direct patient care; so I think the opportunity for AI is to reduce the administrative burden and allow nurses, doctors, and other staff to focus on direct patient care and direct patient interactions—it’s allowing people who chose professions in healthcare to do what they love, and to let automation and AI handle more of the administrative burden. So the lesson is to focus AI use cases to focus on administrative burden.
to focus on administrative burden.