FEATURE · ANALYTICS/AI
Q&A: Blythe Adamson, Ph.D., M.P.H., Flatiron Health’s Senior Principal Scientist
By David Raths O
ver the past several years, Flatiron Health has built up an oncology- focused EHR network and a de-
identified database from more than 280 U.S. academic and community cancer clinics in order to enable large-scale real- world
research.
Blythe Adamson, Ph.D., M.P.H., senior
principal
Blythe Adamson, Ph.D., M.P.H.
scientist at Flatiron, recently spoke with Healthcare Innovation about the company’s laser focus on oncol- ogy data curation and the huge differ- ence machine learn-
ing models have made to the company’s curation and analytical capabilities. According to her bio, Adamson was
formerly the lead data scientist in the West Wing of the White House. She founded Infectious Economics in 2017 to provide thought
leadership to policy makers
and industry leaders on cost-effective strategies to prevent the transmission of viruses. At New York-based Flatiron, which is an independent affiliate of the Roche Group, her team pioneered deep learning language models for extraction of clinical details from EHR documents.
HCI: In 2016, I interviewed Flatiron co-founder Nat Turner about the company’s origins and he told me their initial goal was to create a big-data analytics offering for cancer centers, but then they also realized they had to own the EHR itself, and that became Flatiron’s OncoEMR. Could talk about your role with the company and the increasing importance of AI and the deep learning language models for the extraction of that data from the EHR? Adamson: I've been at Flatiron for five years, and it really has changed over time. The capabilities of what's possible to do now with natural language processing, machine learning, and AI — those capa- bilities didn't exist five years ago when I joined the company or in 2016 when you
first interviewed Nat. What attracted me to Flatiron at that time was that they had solved something that no one in the world had solved before, which was how to curate all of the unstructured data, the clinical notes, the radiology scans. The result of that was this technology-enabled abstrac- tion — opening up these charts in a really standardized way following policies and procedures, and documenting all of those elements of clinical depth that are necessary to be able to get insights that you want about a population or the effectiveness of a treat- ment. It really hasn't been until the last few years, that advances in machine learning have made such huge leaps forward that it's possible to really start taking advantage of that to increase the speed and the scale of curation.
HCI: A lot of people in healthcare have seen this big increase in interest in large language models just in the past six to nine months, but is that something that you guys have been working on for much longer? Adamson: Absolutely. There are so many different types of language models. I put out a paper with my team a few months ago describing at a high level the develop- ment of more than a dozen deep learning models to understand and read language and extract variables with similar accu- racy to what our human clinical experts are doing. The first use case in the cura- tion of EHR data was in cohort selection. We are helping our expert abstractors be more efficient in identifying patients, for example, who might have metastatic breast cancer. But now more of the use cases are actually in training these models to read and identify critical sentences and inter- pret the meaning of sentences in ways that are really similar to these clinical experts following sets of policies and procedures. An example of what that opens up: the bio- marker status for an individual that may be trapped inside a complicated 10-page PDF of a genomic testing report. You might have five different vendors doing genomic sequencing, and all of their reports look different. They're all complicated and dif- ficult to interpret. But one of the advantages
22
hcinnovationgroup.com | JULY/AUGUST 2023
that Flatiron has is a decade of golden data labelled by experts to train these models. That's a really big differentiator because there are a lot of brilliant machine learning engineers all over the world, and a lot of them can design sophisticated architec- tures. But if you don't have the data to train the model, you're not going to be able to get that high-quality performance. For the use cases that that we work with at Flatiron, which can range from commercial insights and academic-level insights to regulatory grade data, we have to keep a very close eye on our approaches for validation and monitoring bias creeping in over time.
HCI: Does the research work you do require everyone to be on Flatiron’s OncoEMR or do you work with oncology practices that might be on some other EHR such as Epic or Cerner? Adamson: We work with both. To be able to answer questions like comparing the effectiveness of two different treatments, we have to have a representative dataset from the U.S., so our datasets reflect 80 percent community oncology care and 20 percent academic hospital systems. Those academic systems may be using Epic or Cerner. We have to do continuous data integration with those hospital systems. And that 80/20 split is on purpose. It really represents where cancer types are seen because there are some cancer types that are only seen in academic hospitals, and there are many for which it's not necessary to go to an academic facility — you can receive high-quality care closer to home.
HCI: Do you think that we'll see the large EHR vendors like Epic and Cerner partner with AI companies to take advantage of these large language models to extract more insights from data and to automate processes fairly soon? Adamson: We are already seeing it! Some of the most common use cases that we're seeing are in the lowest-risk applications. When we think about the benefits and risks of applying AI models, for example, in a clinical decision support tool, there may be different considerations. You would have to
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36
