June 20, 2019
By Julie Gould
Beau Norgeot, MS, of Bakar Computational Health Sciences Institute at the University of California, explains how artificial intelligence (AI) impacts the cost of care for patients with complex diseases, including rheumatoid arthritis, and highlights the challenges of implementing AI systems.
Please tell us a little about yourself and your research interests.
I’m interested in leveraging the data that we are collecting today on individual’s, their health conditions, treatments, and outcomes so that we can make smarter more informed decisions tomorrow. I believe that the best tools to do this at the moment fall into the field of deep learning. So my research is essentially deep learning in personalized medicine with applications in patient similarity, prognosis, and optimal treatment selection. Personally, I have had fairly atypical career path. I dropped out of college the first time and started a few small companies before going back to complete my engineering degrees and PhD.
The findings of your study suggest that models used to forecast complex disease outcomes, like RA, using EHR data is possible. What are the challenges with implementing an artificial intelligence system into practice?
This is a great question with a lot of nuances. Fundamentally, the roadblocks are mostly cultural and policy-driven. Despite being common practice in nearly every other industry, there are both reasonable and unreasonable concerns about the implementation of AI into daily care. As a research community, we need to establish standards that build trust in the dual domains of efficacy and fairness. This will require testing algorithms on larger and more diverse patient populations than we currently see. Our paper was a good step in the right direction, but it’s still not enough. It will also require detailed examination of model performance across subpopulations to ensure efficacy and fairness. Finally, we need a change in the mainstream conversation, aware from replacing doctors and towards enhancing them. Once people feel confident that the algorithms perform consistently and reliably across thousands and thousands of patients, and that doctors will still be driving care, I think we’ll see fairly widespread adoption.
How does the use of an artificial intelligence system impact the cost of care for patients with complex diseases?
AI has the potential to greatly reduce cost of care, especially for patients with complex diseases. Using intelligent systems to get the right treatment to each patient faster could element not only spending on drugs that will not be efficacious for that individual but would also eliminate the need to treat any side effects that may be caused by those drugs. Beyond that, systems to coordinate care, provide virtual medical assistants, and identify clinical trial participants will all help to transform the medical experience and lower the overall cost of care.
How are patient and clinical outcomes impacted by the use of an artificial intelligence system?
This is a really important question, which unfortunately, we don’t have any data points on yet. But the potential for improvement is enormous. Right now, doctors make complex decisions for complex diseases, which are influenced in myriad ways by each patients individual characteristics, essentially through a combination of trial and error and personal experience. It’s easy to imagine that looking at patterns in what does and does not work across hundreds of thousands of patients will lead to improved outcomes across the board.
What are the most important takeaways from your findings? Is there anything else you would like to add?
For me, the vision for the near future of clinical care is pretty clear. Ultimately, we’d like to make medicine more human; to allow machines to do what they do really well so that humans can do what they do really well. Using large-scale real-world data and well-validated AI systems will allow physicians to rapidly select optimal choices for the individual patient in front of them and enable the physician to focus on the human elements of their profession that machines cannot perform; such as uncovering more nuanced symptoms through careful question asking, providing empathy, and removing fear by helping the individual to better understand their situation and options.
The findings from this study are a promising proof-of-concept towards that vision. We’ve shown that AI can perform well on a task that was previously impossible, and that what the machine learned was transferable across vastly different patient populations. Perhaps most excitingly, we accomplished this with a fairly small amount of data, imagine what could be accomplished with nationwide or even global data! Our next step is to do exactly that, to train these models on the national data that we now have access to, and then to launch a trial to evaluate whether physicians armed with these tools can generate better outcomes for their patients than physicians acting alone.
Norgeot B, Glicksberg BS, Trupin L, et al. Assessment of a Deep Learning Model Based on Electronic Health Record Data to Forecast Clinical Outcomes in Patients With Rheumatoid Arthritis [Published online March 15, 2019]. JAMA Netw Open. 2019;2(3):e190606. doi:10.1001/jamanetworkopen.2019.0606