High-risk patient identification at scale: 3 health IT experts answer 5 Qs

Identifying patients at elevated risk for hospitalization before an acute health episode occurs can yield significant benefits for both health systems and patients. The challenge is executing this task efficiently and at scale.


Health at Scale, a San Jose, Calif.-based company led by current and former faculty with ties to Massachusetts Institute of Technology in Boston , Stanford (Calif.) Medicine and University of Michigan in Ann Arbor, among other leading universities, uses machine learning technologies to improve the predictive power of healthcare organizations.

Below, three Health at Scale leaders answer questions about preventative care and the importance of identifying high-risk and rising-risk patients before they become high utilizers. Participants are:

  • Zeeshan Syed, PhD, CEO of Health at Scale
  • Mohammed Saeed, MD, PhD, CMO of Health at Scale, assistant professor and cardiologist at Michigan Medicine
  • Zahoor Elahi, COO of Health at Scale

Question: Why has preventive care and identifying high-risk patients become a bigger priority for healthcare providers in recent years?

Dr. Zeeshan Syed: Preventive care has always been important to providers. What has made it an even bigger priority in recent years is the shift towards value-based care arrangements between payers and providers. With health systems becoming increasingly more accountable for each patient's care over an extended period of time, it has prompted a focus beyond simply reducing the unit prices of services and created a need to get ahead of potentially preventable adverse outcomes. In many ways, we are moving from a short-term cost-based to a long-term outcomes-based set of incentives for providers. Preventive care with early identification of high-risk patients and targeted outreach is a key element in this transition.

Dr. Mohammed Saeed: There are several peer-reviewed studies showing that patients with high past utilization of care services are often on a care trajectory that cannot be reversed. For example, the recent results from the Camden Coalition of Healthcare Providers randomized controlled trial found no difference in readmission rates between patients with high past utilization who received aggressive outreach versus those receiving standard care. It makes sense to try to get ahead of the poor outcomes that underlie persistent high utilization and head them off before they start; this is where preventive care comes in, by treating conditions early so that ballooning costs and condition exacerbation are prevented. To really be effective with preventive care, we need to be able to identify the rising risk patients in a population — patients who will ultimately experience a rise in risk if no action is taken while there is still time and opportunity to impact their care trajectory.

Q: What are some challenges to identifying rising risk patients? Why has conducting this type of outreach proved so difficult in the past?

ZS: It is very difficult to predict which patients are likely to become rising risk at the early stages when their care trajectory is still actionable. These patients can appear at first glance to be relatively healthy and similar to others who will not progress to adverse outcomes. The patterns indicative of future adverse outcomes are often subtle and based on an understanding of changes spanning multiple health parameters over an extended period of time. We are constrained by our limited knowledge of what it means to be at the cusp of worsening health when what it means to be unhealthy hasn't surfaced yet. On top of that, we often have sparse data on patients when they are still rising risk and have not yet progressed to a stage where they may have been extensively worked up with detailed labs and other testing available.

MS: Fundamentally, identifying rising risk patients is a prediction problem. In contrast, identifying high past utilizers is a detection problem. The prediction challenge of finding patients who will be rising risk is much harder, but also potentially much more meaningful and impactful. To be able to predict which patients are rising risk, you need to be able to look at their health longitudinally to assess how they've been progressing over time. Physicians and nurses are trained to be able to do this, and they can, but it requires a lot of time per patient. As panels get bigger, the process to identify at-risk patients gets less and less scalable.

Zahoor Elahi: Outreach has traditionally been done via a "spray and pray" type approach, where care managers have reached out to a broad swath of patients that are already high risk or might become high risk — there is obviously a lot of administrative burden to this approach, making it difficult to spend concentrated time with any one patient enough to make a difference. Another challenge with outreach is that even if you are able to predict the rising risk patients, the current insight doesn't always provide direction as to what to do next to mitigate that risk. You can reach out to the patient and get them in for a visit, but what do you tell them to do, what do you look for? Without knowing the likely cause of their condition exacerbation, it's difficult to determine what to do to prevent it. Current outreach lacks specific next best action for each individual based on their health.

Q: How is technology making it easier for health systems to identify high-risk patients?

ZS: Today we have technology that is able to find markers of future adverse outcomes that are challenging to discover with the human eye and are not currently present in known textbook knowledge but can be identified by machine intelligence when deployed at scale. The patterns and signals identified by these technologies can identify which patients are at high risk in a more personalized, accurate and actionable way. On top of that, we can also apply and deliver these insights in real time; we can adjust our opinion of how to treat a patient with every new piece of data that comes in, as it comes in, to ultimately create a living breathing health monitoring system — which ultimately provides an extremely powerful tool for health systems to take more control of identifying and reaching out to rising risk patients.

Q: What are some barriers to adopting these types of technologies? How can health systems optimize their use?

MS: One of the biggest barriers to adoption is hard-earned skepticism. Health systems have been burned in the past for being adopters of technologies that didn't deliver on promises and created work where it should have made things easier. Not all artificial intelligence and machine learning systems are created equal, and health systems need to get better at recognizing this fact. It should be obvious that like any advanced technology, machine intelligence requires the right expertise, implementation and evaluation to generate better outcomes and return on investment.

ZE: There are a number of ways health systems can approach adopting a machine learning solution to make it more successful.

The first would be to think about doing a pilot to make sure that the technology in question is really equipped to have the intended impact on costs and outcomes. Additionally, health systems should set clear key performance indicators to measure the success of the pilot and implementation. We typically look at total cost of care and reduction in adverse outcomes, such as emergency department visits and hospitalizations, in our deployments to ensure that our solutions are driving meaningful impact.

The second is to be purposeful about operationalization. Insights from machine intelligence must be integrated into existing health system workflows. The algorithms rely on people to act as an effector arm and to take the insights and use them to impact patient care. To be most useful, the machine intelligence must be rolled out in partnership with an effector arm that is prepared and incentivized to utilize the insights from the algorithms.

Q: How do you measure how effective a technology is at identifying rising risk patients?

MS: This is an important question that I think often goes unasked when health systems and other healthcare entities are thinking about bringing machine learning into the fold. There are several ways that you can measure how accurate these technologies are in their predictions. One of the most useful metrics for this kind of problem is the precision or positive predictive value, which is the percentage of patients that are identified as rising risk who then actually have adverse future outcomes. The closer you can get to 100 percent the better. Another key consideration is the timeline of the predictions or how far in advance rising risk patients can be identified with high precision. Being able to predict rising risk 6 to 12 months in advance offers more time to respond and potentially prevent adverse outcomes. Finally, it is also instructive to compare the year-over-year differences in total cost of care and outcomes such as ED visits and hospitalizations for patients identified as rising risk. You would expect that the rising risk patients, if you've predicted that cohort accurately, would experience significant increases in adverse outcomes and total cost of care if their care goes unmanaged.

© Copyright ASC COMMUNICATIONS 2021. Interested in LINKING to or REPRINTING this content? View our policies by clicking here.