November 2019 Edition Vol.11, Issue 11

AI Identifies Patients at High Risk for Unplanned Acute Care

By Megan Garlapow, PhD

The cost of oncology care in the United States has risen since 2010 and is expected to be around $174 billion by next year.1 This perhaps comes as no surprise as the proportion of patients surviving cancer has risen 30% during this time, driven primarily by growth and aging of the population.1  The use of acute care is one critical driver of expanding healthcare spend, with nearly half of all spending for oncology used on acute care.

The American Society of Clinical Oncology (ASCO) described five strategies for decreasing acute care utilization:

  1. Identify patients at high risk for unplanned acute care.
  2. Enhance access and care coordination.
  3. Standardize clinical pathways for symptom management.
  4. Develop new loci for urgent cancer care.
  5. Use early palliative care.2

Interventions may be most effective at reducing costs when targeted specifically to high-risk patients rather than the entire cancer population.2 Identification of high-risk patients therefore becomes a critical priority.

Developing models that detect patients at high-risk for unplanned acute care utilization, however, can be fraught with challenges. In one case-control cohort study at a single center in Massachusetts, patients who received palliative chemotherapy and subsequently experienced unplanned use of acute care had readily available clinical data, such as Charlson comorbidity score and below normal white blood cell and/or platelet count, that associated with chemotherapy-related acute care, but these results remain unvalidated both outside this specific type of cancer care and externally.3

Several studies have identified risk factors associated with experiencing more severe, chemotherapy-related adverse events (grade 3-5), but these studies have been limited to older patients receiving chemotherapy and sometimes further limited to specific tumor types and/or palliative chemotherapy.4-6

Artificial intelligence (AI) harbors the potential to support identification of patients with cancer at high risk of unplanned utilization of acute care. As oncology becomes increasingly value-based, identification of such patients can allow interventions that reduce unplanned acute care use and ultimately reduce costs.

Additional areas where AI could influence value-based care outcomes include identification of patients at high risk for depression, pain, 30-day mortality, and 30-day readmission. Cardinal Health has been working in collaboration with Jvion, a leader in cognitive technology, to develop its Jvion MachineTM  for use in oncology clinics.  The Jvion Machine is an AI tool that identifies patient risk and provides intervention recommendations across multiple use cases or “vectors,” including risk for unplanned acute care use, depression, pain, clinical deterioration, 30-day mortality, and 30-day readmission.7

Amy Valley, PharmD, BCOP, VP Clinical Strategy and Technology  Solutions at Cardinal Health Specialty Solutions, sat with OBR-Green (OBR-G) to discuss the opportunities and obstacles to implementing AI in identification of these high-risk patients. The interview has been edited for clarity and to fit within the scope of this article.

OBR-G: What are the biggest obstacles to uptake of AI in supporting risk assessment?

AV: Some of the biggest obstacles to adoption are infrastructure and physician acceptance and willingness to embrace AI.

Implementation of any sort of new technology, particularly one that is on the front-wave like our AI solution for value-based care, is most successful and achieves benefits quickly when the appropriate infrastructure is in place. As we’re looking at the best candidates for early adopters of AI, though we’re not limiting ourselves to Oncology Care Model (OCM) practices, OCM practices usually have invested in that infrastructure, specifically care-navigation programs and resources. As a result, OCM practices are more likely to be ready to embrace, adopt, and implement our particular AI solution.

Additionally, we have talked to several other practices that did not elect to participate in OCM who also have made transformational changes in how they are managing their patient population. If you have people in place that are focused on value-based care and outcomes for your patient population, then that important obstacle is already eliminated.

The second obstacle to adopting AI is physician resistance, even as physicians become more and more accepting of the concept of AI. We have recent data from our own research at Cardinal Health that shows 53% of oncologists are excited about the role AI will play in supporting care and practice efficiency. Though only 4% of physicians were skeptical of it, it still is a big leap for physicians who have been taught in their medical training that what they are seeing is how they will determine their care plan. With AI and these types of analytics models, physicians are getting a risk assessed for patients and it’s not as tangible as what they might be seeing from other diagnostic tests, such as imaging, etc. It’s a little bit harder to get your head around. Where is the information coming from? How does a machine know that a patient is at risk? But we find that once physicians start using the tool and they experience the machine helping them become more efficient, then they quickly become believers.

OBR-G: How do you overcome that physician hesitance to embrace AI? Is it generational?

AV: Surprisingly, physician hesitance has not been generational. For example, one of our biggest advocates is a medical oncologist who has been in practice for quite some time and has been involved with most of the value-based care models. He participated in several of our panel discussions describing his own experience and journey learning about AI, having some impact in caring for his patients, and then adapting what he had learned into care management in his practice. Interestingly, some of the other physicians in his practice had a little more hesitancy.

We hear similar stories from other practices, so I don’t think it’s fair to assume that hesitance will be generationally stratified. Oncologists require documentation and research at the same level of rigor as we would see in a clinical trial for a drug. Then, oncologists will consider these AI solutions and how they can adopt AI into patient care.

OBR-G: How do you promote adoption of AI in the community setting? How is development of AI in the community setting different than in the hospital setting?

AV: The community oncology world is very engaged and actively participating in value-based care. Increasingly, evidence indicates that care in the community setting is of the highest quality and lowest cost. In addition to being engaged in value-based care, community oncology practices are developing and using tools to manage their patient population. As a result, community oncologists are interested in ensuring acute care utilization decreases as a way to demonstrate cost savings and value to payers. Community oncology practices are just as capable and sometimes nimbler than academic medical centers at implementing new technologies.

Across all of healthcare, decreasing acute care utilization has become increasingly important and community practices are putting great efforts into keeping patients from unnecessary emergency room visits or hospitalizations.

Our AI solution is something that a community oncology practice could use, and hospitals with outpatient cancer clinics could also benefit from this solution. Everyone is trying to perform better in value-based care, and we’re trying to provide better quality care for our patients while we decrease total cost.

OBR-G: At what threshold of accuracy of stratifying patients into risk categories for utilizing acute care do AI approaches become worth it? How do you balance false positives and false negatives with the overall cost savings to the healthcare system?

AV: This is a very important question because we don’t want to just apply artificial intelligence without understanding if we really are making a difference. And if we really are making a difference, we have to balance out false positives and false negatives.

One of the first things Cardinal Health did when we turned on the Jvion Machine in the pilot practices is validate that the patients identified as high-risk were concordant with current manual processes of high-risk identification at the practices. And we found a high degree of concordance in that validation cohort.

We don’t know the exact threshold of accuracy that we’re going to accept, but we work within an acceptable range to proceed with using the machine in those practices. We still need continued experience and documentation of the outcomes to determine that threshold of accuracy. What we’re seeing now is a tremendous opportunity to bring efficiency to how we identify high-risk patients, which enables us to put them through the appropriate triage pathways to try to reduce the risk for acute care utilization; to reconsider care plans; or to provide referrals to pain management, mental health professionals in the case of depression risk, or to palliative care and hospice as appropriate.

We are working with two of our pilot practices to design a prospective trial assessing the impact of the Jvion Machine. We have continued to evaluate the outcomes of the patients that we’re identifying against performance from several OCM metrics over time. As the results mature, we’re seeing a continued and increasing benefit in all of the endpoints, whether that’s predicting patients at high risk for 30-day and 90-day mortality or looking at patients at risk for pain, for depression, or for clinical deterioration in six months. All of those metrics continue to improve. We need to have the same rigor we would have when evaluating any new therapy or any new procedure for patient care, so we’re proceeding with prospective clinical trials.

OBR-G: How does AI need to be developed differently for use in clinical decision making in general acute care vs end-of-life and palliative care? Why?

AV: The physician is at the center of clinical decision making. This is critical. We don’t envision AI replacing that physician, that clinical judgment, and really the holistic view of the patient that the physician has. For example, when the machine identifies a patient who is at risk for 30-day and 90-day mortality, that is just one piece of information that the physician is going to consider when evaluating that patient’s care plan.

Decisions physicians make regarding care plans include consideration of information from the AI machine. But I don’t foresee an instance where the AI machine indicates a patient is at risk and then is removed from active treatment without further consideration and placed in an end-of-life pathway. Instead, our physicians are viewing results from the machine as another lab test or another piece of information to consider in the total view of the patient.

OBR-G: What results from working with the Jvion Machine are you most excited about?

AV: I am most excited about the ability to make an impact very quickly.  When we get the Jvion Machine engaged with a patient, some time is needed for the data to be integrated and normalized, but within 12 to 16 weeks, we’re starting to see tremendous impact.

When we launch with a practice, we start with one vector at a time, such as risk for 30-day mortality. When the practice has adapted that into their workflow, they’re comfortable using the portal, and we’ve overcome some of those earlier hesitancies around using AI, then we can go to the next vector, such as risk for pain or risk for depression. Each of the next vectors come in very rapid succession. It’s not very long between when you get the data flowing through to when clinicians and care navigators can start seeing the benefit from using the machine.

The other part that is so exciting are the stories we hear about patients identified as high-risk that physicians didn’t realize were at risk. The machine pulls in social and behavioral determinants of health that are sometimes so difficult to evaluate in conjunction with all of the clinical data. Use of the Jvion Machine is making an impact in people’s lives, doing so in a very meaningful way, and doing so fairly quickly. That’s what excites us as we’re developing these types of solutions that benefit not only the community oncology practices but also the patients that they care for.

OBR-G: Where do you see the Jvion machine going in five years? What do you think the landscape will look like in a decade?

AV: In five years, there will be even more acceptance of AI. As acceptance continues to grow, we’ll really be reaping efficiencies in how we navigate these massive amounts of information that are coming to us from new technologies, new therapeutics, and new knowledge about cancer management. AI is going to play a huge role in sifting through all of these data and in helping clinicians make tough decisions, design care plans, and help patients get the best outcomes for their treatment. I think AI is going to become much more mainstream in 5 years. Enabling faster implementation of science to the bedside and being able to navigate today’s operational challenges are some of our goals for working with the Jvion Machine.

I don’t even know what to expect 10 years from now. Who knows what healthcare will be like in 10 years? The pace is so rapid in terms of advancing science and the technology. Our current healthcare is based on a model where people travel to see their physician and we engage them. In 10 years, I think the technology is going to connect experts and care providers with patients in an entirely different way, in their homes and in a more global way. In my ideal situation, in 10 years, I could be anywhere in the world and be getting my cancer care and have more access to rapidly evolving science and the experts I need to get the best outcomes as a cancer patient.


  1. National Cancer Institute. Cancer prevalence and cost of care projections. Accessed October 8, 2019.
  2. Handley NR, Schuchter LM, Bekelman JE. Best practices for reducing unplanned acute care for patients with cancer. J Oncol Pract. 2018;14(5):306-313.
  3. Brooks GA, Kansagra AJ, Rao SR, Weitzman JI, Linden EA, Jacobson JO. A clinical prediction model to assess risk for chemotherapy-related hospitalization in patients initiating palliative chemotherapy. JAMA Oncol. 2015;1(4):441-7.
  4. Hurria A, Togawa K, Mohile SG, et al. Predicting chemotherapy toxicity in older adults with cancer: a prospective multicenter study. J Clin Oncol. 2011;29(25):3457-65.
  5. Extermann M, Boler I, Reich RR, et al. Predicting the risk of chemotherapy toxicity in older patients: the Chemotherapy Risk Assessment Scale for High-Age Patients (CRASH) score. Cancer. 2012;118(13):3377-86.
  6. Retornaz F, Guillem O, Rousseau F, et al. Predicting chemotherapy toxicity and death in older adults with colon cancer: results of MOST study. Oncologist. 2019; doi:10.1634/theoncologist.2019-0241. [Epub ahead of print].
  7. The Jvion Machine: AI that works for oncology. Accessed October 11, 2019.


About Cardinal Health

Cardinal Health Specialty Solutions empowers community oncology practices to navigate the future of oncology care. Working in close collaboration with each practice, we develop and implement a clear roadmap to meet pressing challenges in a new era. With expert insights, responsive tools and committed support through VitalSource™ GPO, we enable care providers to deliver high-quality, cost-efficient care and future-proof their practices. Discuss your roadmap today.

Post a Comment

OBR Archives

To view previous issues of OBR green you can visit our archives. The entire library of OBR green articles is searchable.