The idea that artificial intelligence (AI) can be an effective tool in diagnosing cancer has come a long way, particularly since the once hyped-up project between IBM Watson and MD Anderson Cancer Center ended as a $62 million failure. But there is growing support that AI could possess the ability to recognize patterns that can be too subtle for the human eye. Now, oncologists are increasingly using AI to look beyond the patient’s clinical state to understand their social, behavioral and environmental context.
There are many social determinants of health (SDOH) that factor into a patient’s health, and for cancer patients specifically, oncologists are paying more attention to key environmental considerations, such as whether a cancer patient has transportation to their appointments, can afford their treatment, or has a strong support system. These are powerful determinants of outcomes, but are often invisible to oncologists and overlooked in care plans, for a variety of reasons. For instance, in a 15-minute physician-patient interaction, because there are so many important details related to treatment, side effects, and reviewing recent scans of what the cancer cells look like, there simply isn’t an opportunity to ask patients about what else is going on in their lives from an environmental or behavioral perspective, contends Ray Page, D.O., Ph.D., lead oncologist at the Center for Cancer and Blood Disorders, a Fort Worth, Tx.-based cancer treatment center serving more than 12,000 patients across its nine clinic sites.
In a recent interview with Healthcare Innovation, Page notes the massive importance of using technology such as AI to comb through external data and bring SDOH factors into the clinical workflow so that providers could address patients’ non-clinical risk factors in their care plans. The Center for Cancer and Blood Disorders is specifically leveraging an AI platform from Jvion, a developer of analytics software, that looks at thousands of clinical and non-clinical factors per patient to provide oncologists with a better understanding of their patients’ hidden risk factors. It then recommends interventions that address each patient’s specific risk factors, according to officials.
In this same context, one of Jvion’s partners, healthcare solutions company Cardinal Health, spearheaded a study earlier this year of 160 oncologists; 90 percent of respondents said they believe SDOH factors such as financial security, access to food and social isolation are significantly impacting outcomes for cancer patients. That research further revealed that 68 percent of the participating oncologists said at least half of their patients are negatively impacted by SDOH. Financial insecurity/lack of health insurance (83 percent) was by far the most cited barrier for patients according to participating oncologists, with access to transportation (58 percent) and health literacy (53 percent) also among the top responses.
In the interview, Page discussed the importance of incorporating SDOH data for cancer patients, how his organization has progressed in this area, and how AI is helping the process. Below are excerpts of that conversation.
Looking back, what was your team’s motivation to leverage AI to look beyond cancer patients’ clinical states?
There are lots of payment reform models out there, and when the MACRA [Medicare Access and CHIP Reauthorization Act of 2015] legislation came about, there was a big push toward transitioning from fee-for-service payments to payments being made for the quality and comprehensive management of patients. One such alternative payment model, called the Oncology Care Model, or OCM, has [nearly 200] entities or practices participating in it. [The model, from CMS’ Innovation Center (CMMI), includes physician practices that have entered into payment arrangements that include financial and performance accountability for episodes of care surrounding chemotherapy administration to cancer patients].
Our practice was one of those that enrolled into the OCM alternative payment model, and we realized pretty quickly that the methodology, and the payment methodology, were extremely complex. We were basically responsible for the total comprehensive management of the cancer patient. So it went way beyond fee-for-service and involved having some sort of control over the entire comprehensive care of the cancer patient. When we looked at that, and looked at the payment model with our large patient population that we have in the North Texas area with nine clinic sites, we realized that we were going to need to do some risk management, and to try to identify our patients who are at the highest risk of having adverse outcomes based on their social determinants of health.
So I sat down with our nurse manager and we started to come up with 25 different variables/parameters that we thought would identify high-risk patients, and from there we could get physicians and case managers on to do triage management for those individuals. As I was in the process of pulling together those 25 variables, I found out very quickly that doing this homegrown in my own practice was quite an effort. At [that same time], I was at a national meeting and saw a presentation on Jvion being explored in oncology. I was blown away with the presentation that they had, and became extremely interested in getting it engaged in our practice, and having the ability to use artificial intelligence to do a better job than an individual can in identifying those patients and risk stratifying them. So our practice joined together with a couple of other practices across the U.S. to pilot the development of an AI-based risk stratification tool to help us best identify our patients who are at the highest risk of having adverse outcomes. We plugged that into our system and worked closely with Jvion to power that AI-based tool.
What were the variables that you ultimately decided on?
One example is looking at who within the next 30 days might have a pain management issue where they need to be put on opioids or see a pain doctor? Also, how many of those [patients] are going to face depression, and will need antidepressants or psychologic services over the next three to six months? Who is going to have a deterioration over the next six months? Who can we avoid sending to the emergency room or hospital? Who is in the hospital that we can predict is going to beat high risk of having a readmission? Who do we anticipate could potentially die within the next 30 days, where we need to be more aggressive about incorporating palliative care services?
So we looked at seven vectors with Jvion to try to identify the risks of those patients. The cool thing about the Jvion machine is that while I was trying to come up with 25 features to identify patients, Jvion looks at over 4,000 variables. And those variables come from everywhere: they come our EHR, from hospitals, from payers, and from data that's just out there in the community. The [platform] can explore thousands of variables, including all the social determinants of health. They were [ultimately] able to take all of our patients that we have in our [organization], look at the different vectors, and actually rank patients and stratify them based on what shows up as red flags on their algorithm. And then they give us reasons why somebody may be a risk of having a certain problem.
From there, we can get a dashboard and a report that we analyze weekly, and work on with our physicians and our case managers. For those people who are at high risk of having one adverse event or another, we can plug in a variety of service lines that we have. So maybe somebody needs to see a pain doctor or palliative care doctor, or maybe they need rehabilitation or to see a dietician. We can incorporate all that and improve the outcomes for our patients, and therefore improve our performance, particularly when looking at payment models like OCM.
Did you find that it was the SDOH data that was really able to help you risk stratify and give you the needed predictive element, or was it a true mix of clinical and social data?
It’s a mix. They look at thousands of variables and all those things come into play. They do give us reports where they figure out the leading variables; whether or not patients live in a poor zip code, whether or not they own a house or rent an apartment, or whether or not they lost a spouse or have a spouse with an illness. Then there are the number of doctors’ visits, access to food, and other parameters centered around debility, etc. So there are many variables that get funneled down and distilled into ranking or categorizing that patient.
In oncology specifically, has there historically been a gap in learning about and integrating SDOH data?
Obviously with cancer patients, they are some of the sickest patients you have. When you get cancer, you can lose your job, your insurance, and can run into financial issues. In the world of social determinants of health, cancer patients are extremely complex. And yes, as far as cancer doctors and cancer centers go, we have done an extremely poor job of identifying those [SDOH] in the past.
In a busy oncology practice, I might see 25 patients per day in the clinic. So, during that time, the doctor may have a 10-minute interaction with the patient, but that interaction is focused on a variety of things: the physical examination, talking about the cancer, reviewing the patient’s scans, talking about the pathology, about the type of chemotherapy and side effects, and all those kind of things. So you have 10 minutes to get to all of that clinical information that’s pertinent to administration and management. It is absolutely impossible for us to sit down and start addressing some of the issues with the patient such as how things are going with your nutrition, your transportation, your finances, your spouse and your job. There’s just no way that you can gain that wealth of information in that clinic visit. We need other resources, and this has been a good tool to be able to explore that as a resource to identify patients who are in need.
What are some quantifiable results you have been able to see from this initiative so far?
We do have some information that has been published in conjunction with one of our sister practices in Tacoma, Wash. Between our two practices, when we looked at all the OCM data, and the Medicare payments we had a substantial reduction in ER visits and hospitalizations. We presented these findings at the 2019 Community Oncology Alliance (COA) conference. Just in Medicare alone, we saved over $3.1 million over a six-month period. And that was in the context of OCM and using simple management and case management protocols, but also using the AI tool to identify those patients. The Tacoma practice also had a 225 percent increase in hospice referrals and a 35 percent increase in palliative care referrals.
