The Cleveland Clinic is no stranger to artificial intelligence.
The large health system launched its Center for Clinical Artificial Intelligence in 2019. It announced a 10-year partnership with IBM last year to accelerate the use of AI, hybrid cloud and quantum computing technologies. And in January, the company hired its first-ever chief digital officer to oversee the clinic's digital transformation, including in areas such as artificial intelligence, machine learning and big data.
Most recently, Cleveland Clinic announced a five-year partnership with PathAI, a Boston-based company that uses machine learning to identify pathological predictors of the effect of prescription drugs. As part of the deal, Cleveland Clinic will become an equity holder in PathAI.
“We don’t have a lot of AI and machine learning in my institute,” says Dr. Brian Rubin, chair of Pathology and Laboratory Medicine Institute at Cleveland Clinic. “I thought, ‘How can we take advantage of the recent developments in pathology technology? Looking around, the idea of hiring people and building this piecemeal seemed daunting. So, I approached PathAI with the possibility of forming a partnership. Building this on our own would have taken a lot of time.”
Digital Health Business & Technology spoke with Rubin about the potential of AI to improve pathology and lab medicine, how he plans to learn from high-profile health system-AI failures and more. This interview has been edited for length and clarity.
What potential do you see for AI improving pathology and lab medicine?
Pathology is a subjective area and looking at glass slides is a subjective process. Most people agree on most diagnoses, but the stakes are very high. When you make a diagnosis of breast cancer, a portion of a breast or an entire breast is going to be surgically removed from a woman. Pathologists don’t take that responsibility lightly. We understand the consequences of our diagnoses. There are a lot of routine cases, but a small percentage of cases are very difficult.
I think that AI-assisted diagnosis will bring a reproducibility and a quality that only a computer can really bring for those difficult cases. It’s going to do things the same way every time. There’s a partnership between the pathologist and the algorithm to do that. I think that there's also a very tiny amount of pathology that can be missed. One of the early benefits will be algorithms that can scan through all of our glass slides and identify things we may have missed so that we can look at them, and then decide if it's a true miss or it's just something that's not clinically important.
One of the most important things we do with a microscope almost every day is to grade cancer because that grading dictates therapy. A high-graded cancer would receive chemotherapy, which is very expensive and there a lot of side effects. It’s not something that we take lightly. If a patient needs chemotherapy, we want to give it to the right patient. And a lot of grading is based on counting of mitotic figures. Computers are great at doing that. Why not leverage the strength of the computer to count mitotic figures?
Download Modern Healthcare’s app to stay informed when industry news breaks.
How do you measure the success of this collaboration from a clinical perspective?
One of the things we want to do is build our own internal group. They're going to do a lot of teaching and educational activities—teaching our residents, teaching our staff, etc. Building our own internal capabilities to do algorithm development will be one measurement of success. At the end of the day, adoption of algorithms by pathologists is going to be the measure of success here. And from there, we can run prospective clinical trials, measuring AI-assisted diagnosis versus traditional diagnosis and whether the outcomes benefit patients. Prospective trials will be a measure of success. It is going to be a slow process. It's not going to happen overnight.
How do you overcome some of the past challenges that have hurt other health system-AI company partnerships?
There have been a lot of high-profile failures. What I think is most important in pathology, and it's probably equally as important in every other area of medicine, is having a lot of domain expertise when developing any of these tools. Pathologists have to be integrally involved. They must be an important part of the DNA of anything that's being developed clinically for their field. If I had to guess what’s failed in these other partnerships, it’s they became very computer and AI centric. There was potentially an arrogance that computer scientists can solve all these problems. That’s probably why these things fail. The people who use these tools have to be integrally involved every step of the way. That’s why this partnership works for me because we're going to be heavily involved in every part of the process of developing algorithms.
Are you concerned about any potential data biases in algorithms?
One of the things that’s nice about the lab is that we know how to validate testing. We approach AI algorithms like a laboratory-developed test. It requires a ton a validation. We won’t ever use a test that we're not convinced has been validated at every possible level. That means a lot of validation. I think prospective clinical trials will probably be the best way to go out and see how these things test. We’re going to use it in a variety of environments. We want to pressure test these algorithms in multiple different settings. I think rigorous testing, validation and prospective clinical trials all have to be part of the equation.
