Evaluating AI for Academic and Private Healthcare Settings
Artificial intelligence (AI) is becoming widely available in hospital radiology departments and standalone imaging centers to help detect diseases earlier, standardize imaging data analysis, increase access to care, and reduce the burden of repetitive tasks on radiologists. But as the quantity and types of AI solutions continue to grow, private and academic healthcare centers will require different approaches to assess how to successfully incorporate AI into clinical practice.
To shed light on this topic, Applied Radiology hosted a recent roundtable discussion evaluating the differences in implementing AI in academic and private settings. Sponsored by Guerbet, this was the third segment of AI Insights, a series of webinars focusing on the growing use of AI in healthcare.
The three-person panel was led by Christopher Filippi, MD, radiologist-in-chief at Tufts Medical Center in Boston; it also included Lawrence Tanenbaum, MD, FACR, vice president, chief technology officer and medical director for the eastern region at RadNet in New York City and a member of the editorial advisory board of Applied Radiology; and Aashim Bhatia, MD, MS, a pediatric neuroradiologist at the Children's Hospital of Philadelphia (CHOP).
Similar goals for all facilities
As imaging volume increases across settings, Dr Tanenbaum said there’s a “mismatch between the volume we have and the manpower we can recruit,” but AI can help bridge the gap to improve efficiency and reduce burnout.
“The types of AI that an outpatient organization is interested in is very different from the ones used most commonly in the hospital. But ultimately there is a thread behind all of them, which is workflow efficiency,” he explained.
Private and academic settings will benefit from the promise of AI, said Dr Filippi. “I want AI to fully automate the mundane tasks that are leading to burnout in radiologists. It should automatically measure and segment without us having to do it. It should automatically protocol and schedule patients to make our lives more efficient and better.
However, Dr Filippi said, efficiency in an academic setting like a teaching hospital might look different from that of a private setting such as a community hospital. Most academic centers have experienced radiologists who can easily identify a hemorrhage, large vessel occlusion, fracture, or pulmonary embolism. A smaller hospital might need AI to help detect those pathologies, while the academic center might want AI primarily to reduce the burden on radiologists reading scans 24/7. “AI might help our residents who take call at night make less mistakes,” explained Dr Filippi.
Evaluating the AI need
The panel agreed that the biggest challenge in evaluating AI is filtering through all the options and deciding what will work best in a specific facility. “I can only get traction in my organization for things that align with my outpatient organization's goals, whereas academia has the ability to dream. And that's a big difference in terms of evaluating which ones you get to use,” said Dr Tanenbaum.
Dr Bhatia noted that he also has to consider AI-based solutions that will impact the entire organization. “It comes back to what software is best for the whole. If I want AI for neuroimaging, but it won’t affect the majority of the radiologists in this hospital, then that's probably not the software the department's going to invest in initially,” Dr Bhatia said.
Dr Tanenbaum said private imaging clinics and hospitals are often more focused on improving everyday workflow. “We don't have the high-minded goals of academia. But we do have the same fundamental requirements that it elevate our performance, improve our efficiency, and be something that we believe in.”
Dr Filippi agreed, noting that reimbursement continues to pose challenges across settings. He believes radiologists should work with industry associations and payers to resolve reimbursement issues around AI. “It's important for radiologists to band together, have a united front, and to get involved with things like [the] ACR. I get involved with my local insurance providers all the time to get reimbursement,” he said.
Opportunities in academic centers
One of the biggest advantages of academic settings is their ability to develop AI internally. During his fellowship at a teaching hospital, Dr Bhatia wanted to develop an algorithm for stroke detection in partnership with an AI vendor. While there were benefits to doing this, he also ran into significant roadblocks.
“The advantage of a large academic center is the large data sets to train AI algorithms. But implementing it is the biggest challenge because you have security and data storage issues. It’s also unclear what access the vendor will have to your patient data. And after the trial, [what is the cost] of using that software?” said Dr Bhatia.
Another significant opportunity for academic centers is radiogenomics, which combines quantitative data from images with individual genomic phenotypes to create a deep-learning prediction model that stratifies patients, guides treatment and evaluates outcomes. Dr Tanenbaum said it’s impossible to develop a radiogenomics program in a private practice owing to the need for large data sets and other resources only found in teaching hospitals.
Dr Bhatia said this is particularly important in pediatrics, where there's more diversity than in adult patients. “Radiogenomics at academic institutions like CHOP are possible because you have a large dataset and you’re able to train datasets to differentiate tumors. These statistical methods look promising, but they have to be tested prospectively in an academic center,” he explained. Dr Filippi agreed, saying “When we try to translate these algorithms outside the academic center, to the private sector, it doesn’t work so well.”
This inability to translate internally built AI to the private sector sometimes lies within the data itself. “For these algorithms, you need really large datasets. And the problem is that these datasets come from historical data, and some of this historical data comes from institutions where there’s been a lot of inequity and bias against certain groups of people,” said Dr Filippi. “I’m always concerned when I’m evaluating algorithms, whether it’s in my academic practice or in the community hospitals that are part of our network. Is it matching the demographic of people we serve? If it’s not, it’s not likely to work.”
The benefits and challenges of private settings
In private imaging clinics and medical centers, radiologists are often challenged with reading large numbers of cases. AI could screen these cases for acute abnormalities, which would also improve outcomes. One area that’s essential to outpatient groups like RadNet is the need for standardization.
“If I have 75 neuroradiologists, there is no standardization between their reports, which diminishes the value of what we can deliver,” Dr Tanenbaum said. “AI tools that appeal to us are ones that improve the standardization of our output, so our brand becomes something that you can rely upon. They know a RadNet case is going to be read in a consistent way no matter who is reading the study.”
Wide-ranging benefits for all settings
Regardless of setting, private or academic, the panel agreed, AI will have a long-lasting impact on radiology. “It’s to our detriment if we don't use AI to make our lives more efficient and better,” said Dr Filippi. “The big promise is that we can use AI to move toward precision health, to improve outcomes and reduce healthcare inequity.”