The company wants to make family caregiving coordination as common as paternity and maternity leaveRead more...
The technology can automatically write parts of radiology reports, saving radiologists from burnout
Radiologists, like so many in the medical field, are experiencing burn-out at an alarming rate. According to a report from Medscape from earlier this year, radiologists have a burn-out rate of 44 percent, with another 14 percent saying they are depressed, numbers that are on the same level as other practicing physicians. Nearly half the radiologists in the study said they spend too much time at work.
That's why Dr. Jeff Chang, a practicing radiologist, founded Rad AI, a startup that seeks to make the radiologist more efficient by using machine learning and artificial intelligence to automate the most repetitive parts of their job.
"The amount of imaging being done goes up by about two or three percent per year, while the number of radiologists in the U.S. stays flat, or goes down slightly. Also, CMS, and other private insurance, reimbursements go down every year, so there’s really this question of how to use technology to become more efficient, in order to continue to allow radiologists to do what they do best," he told me in an interview.
The company, which was founded in February of last year, officially launched on Monday, along with a $4 million seed round led by Gradient Ventures, Google’s AI-focused venture fund. UP2398, Precursor Ventures, GMO Venture Partners, Array Ventures, Hike Ventures, Fifty Years VC and various angels also participated in this round, which is the company's first funding.
Rad AI's first product automates the impression section of a radiologist's report, which Chang described as both an integral, yet also repetitive, part of the process; one that, if done by a machine, will free up the radiologist to get more work done in the same amount of time.
"The radiology report contains a number of different sections. The findings are where you describe all the things that you see on the images, and then the impression is where you both summarize and synthesize the conclusions of what you’ve seen, and also the appropriate follow-up recommendations. Based on past research, it seems that only about 50 percent of ordering clinicians read anything besides the impression, so the impression is considered the key part of the report," he explained.
"At the same time, it is sort of repetitive, you’re summarizing and synthesizing conclusions based what you said in the findings. So, essentially, what we do is automatically generate the impressions based on the rest of the report, the findings and the indication."
Rad AI's product can even write in each radiologist’s preferred language, meaning that, based on how they use language and their separate model encoding, it automatically generates language that sounds like them. The radiologist can then just review the impression section, and then sign off.
With this new capital, Rad AI says it will build out its engineering team; the company, which currently employs 13 people, with eight engineers, plans to grow to 30 employees in the next year, around 20 of which will be engineers.
Rad AI, which works directly with private radiology groups, as they account for between 70 and 80 percent of the market, also plans to also use the funding to expand to more partners in the coming year.
Rad AI is currently working closely with five groups, including Greensboro Radiology, Medford Radiology, Einstein Healthcare Network, and Bay Imaging Consultants, one of the largest private radiology groups in the United States. It has over two dozen partnerships that are currently being negotiated, so the plan is to have 15 to 20 partners by the end of 2020.
Deploying machine learning in the radiology space
Until now, radiology wasn't a field that typically implemented technologies like artificial intelligence and machine learning, and for good reason, according to Chang: they didn't actually make the radiologist's life easier. In fact, he told me, many solutions made it more difficult.
"It’s really difficult for machine learning companies to work in radiology because the radiologist already has such a well-tuned workflow. So, for any product to actually add to radiologist’s efficiency or well-being is really difficult. Usually, when a team comes in from a deep learning or engineering perspective, they find a problem that’s easily tractable that they can solve with a model, like, ‘Can we identify this finding on an image?’ but the radiologist can already do that, and 100 other findings, on that image, so it doesn't add significant value," he said.
"So, it’s really difficult to find initial products that radiologists actually want to use. When you ask at conferences which radiology groups are using AI products, it’s less than 5 percent, even though a number of these companies have been around for almost a decade."
Part of the difference with Rad AI is that Chang, as opposed to other founders in this space, actually comes from the world of radiology; he was the youngest radiologist, and second youngest doctor on record, in the US, having gone to college at 13 and med school at 16. He wound up becoming a radiologist, he said, "mainly because I was particularly interested in the intersection of technology and medicine. Radiology is heavily dependent on advances in technology, and is among the fastest-changing fields in medicine due to technological trends."
The other thing that separates Rad AI from other companies in the space is that it is all about making the life of the radiologist more efficient by taking a part of their workflow out, rather than either adding more steps in, or just prioritizing which steps to take first.
"Because I’m coming in with a background of having been a practicing radiologist for the past decade, I really understand the radiology workflow. When we talk about radiology workflow, we’re talking about PACS, voice recognition, RIS, a smart worklist – all these things have been built out for radiologists to be as efficient as possible. The issue with many startup’s products is they disrupt that workflow, that they add additional steps, additional images, for the radiologist to look at. That additional work slows the radiologist down," said Chang.
"Given that we, as radiologists, need to read 100 to 200 studies a day, it’s really hard to add anything into your workflow that slows you down. So, the key for us is to make sure, as we roll out our product, that it makes you more efficient, while maintaining or improving clinical quality."
Even a company like Aidoc, which uses deep learning algorithms to analyze medical imaging, allowing radiologists to better detect abnormalities, and which raised a $27 million round of funding earlier this year, only reprioritizes certain reports, according to Chang; it doesn't actually reduce any of the work the radiologist has to do.
"They put a study higher on the worklist, so that it gets read before other studies, but, overall, it does not change productivity or efficiency, it’s changing the order of the studies," said Chang. "They’ve been working with several academic centers to see whether it actually makes a difference in turnaround times, but so far, because of the amount of underlying noise in how quickly things get done, it’s been hard to show a significant change in turnaround times."
Rad AI, meanwhile, has already seen major time savings for radiologists; freeing up nearly a quarter for their time to do their other work, meaning they can get through their studies faster.
"Radiologists, typically, on any particular day or shift, read anywhere between 100 and 200 studies, so there’s almost always going to be more work that you can catch up on. There’s always a worklist where more studies are being done, so this helps radiologists not fall behind on their worklist each day," said Chang.
Augmenting the radiologist
Going forward, Rad AI plans to expand from automating reports to incorporating the image workflow, and it has a number of products in the pipeline to increase radiologists’ time savings and efficiency.
"We have a number of products in the pipeline, focused on: how do you improve radiologists’ efficiency further, while reducing radiologist burnout? How do you automate certain manual repetitive tasks that radiologists are currently doing? Much of that will gradually include a combination of both imaging detection/diagnosis and the report," said Chang.
"We have a really deep understanding of radiology reports and one of the larger datasets in the U.S. for radiology reports, and we’ve also been building a similar dataset for images, so our next products will involve how to automate more parts of the workflow for radiologists, and how to save them more time."
However, he stresses that his technology is not going to replace the radiologist; rather, it will make them better at their jobs. And, he noted, there will always be a human in the mix to ensure everything is done correctly.
"You always want to have radiologists looking at the images, reviewing whatever the machine learning algorithm is generating. You’re simply making it faster for radiologists. So, instead of having to manually dictate the entire report, most of it might come up automatically, and they can still review, they can determine whether it fits their language, they can determine if there are additional things that need to be added. So, the question we’re constantly asking ourselves is, how do you provide greater automation for radiologists so that they can save time?" he said.
In fact, five years from now, Chang believes the job of the radiologist will be basically the same, as will much of the current workflow. The basic fundamentals of radiology will remain the same.
"It’s just that certain parts of the workflow can be completed much faster. So, the number of words radiologists have to dictate might drop by 60 or 70 percent, because a lot of it is already filled out, so they can review whether it’s what they would have said, or if they need to change the language. There’s a lot that you can automate, in terms of helping the radiologist not have to do as much manual work. At the same time, it’s always up to the radiologist to determine whether this is an accurate result. Radiologists have the final word on the reports that they provide to their ordering clinicians and patients."
Support VatorNews by Donating
Read more from our "Trends and news" series
Digital health news, funding roundup in the prior week; May 16, 2022Read more...
The platform gives them access to tools around community, education, technology, and well-beingRead more...
Related Companies, Investors, and Entrepreneurs
Joined Vator on
Rad AI applies neural networks to radiology in order to save lives and reduce the cost of healthcare.
Our first product integrates seamlessly with existing workflow to significantly increase radiologist efficiency, while reducing radiologist burnout and improving clinical accuracy.