How to predict which patients will be hospitalized

What’s the most effective way to predict which patients will be hospitalized in the coming year? The Heritage Provider Network, a managed care group in California, hopes to answer this question through its sponsorship of the Heritage Health Prize, a $3 million X Prize-like competition for health care. The contest invites participants to develop a prediction algorithm to identify patients who will spend time in the hospital in the following year and check their methods against three years of Heritage’s own claims data. S/he who most accurately predicts hospitalizations takes the prize.

Heritage Network will benefit directly from the results, as will other insurers facing skyrocketing medical costs. Unnecessary hospitalizations account for billions in excess health care spending, so there is a tangible financial incentive to predict who is likely to be hospitalized, and for how long. Identifying patients most at risk for hospitalization, and preventing or reducing the length of these hospitalizations, will allow providers to “develop new care plans and strategies to reach patients before emergencies occur,” according to the contest’s official description.

The goals of this contest are admirable, and could lead to important improvements in America’s health care delivery system; however, the dataset is missing critical information about the single biggest driver of health care spending: geography. The data include information on patients’ age, sex, diagnosis codes, and previous length(s) of stay, but excludes the name and the location of the hospitals where patients received care. According to a forum on the Heritage Prize website, geographic data were withheld to protect patient privacy, but by doing so contest organizers ignore the unavoidable, albeit unpalatable, fact that both quality and practice habits in U.S. health care vary widely from region to region and hospital to hospital.

Research by the Dartmouth Atlas Project has repeatedly demonstrated that there is significant variation between hospitals when it comes to length of hospitalization and the intensity of physician services provided to medically and demographically similar patients. Dartmouth researchers have even quantified this variation in a statistic known as the Hospital Care Intensity (HCI) index, which allows for a comparison of intensity levels among hospitals across the country. Growing acceptance of the geographic underpinning of variation in health care delivery has led to many regionally specific quality improvement efforts across the country. A recent Robert Wood Johnson Foundation Alliance for Health Reform event highlighted some of the best of these projects, which seek to connect multiple stakeholders in a community in an effort to tailor best practices of the delivery system to a region’s specific needs and characteristics.

These strategies could resemble the approach taken by Dr. Jeffrey Brenner and the Camden Coalition of Healthcare Providers to target the “super-utilizers” of emergency services, as described in Atul Gawande’s article in the January 24, 2011 issue of The New Yorker. According to Gawande, one percent of the patient population in Camden is responsible for nearly one-third of all medical costs. An algorithm to accurately identify potential “super-utilizers” would be a valuable asset to the Camden Coalition, Heritage Health Network, or any other insurer, and could lead to significant cost savings. The application of such an algorithm across the entire population has the potential to address the nationwide problem of skyrocketing health care costs.

Given the existence of variation in the intensity of medical care, knowledge of which hospital a patient is treated at could greatly improve any algorithm’s ability to predict the frequency and duration of hospitalization. The organizers of the Heritage Health Prize would do well to release geographic data to contest participants to allow for the development of a more powerful predictive tool that accurately reflects the reality of variation in American health care delivery. That would be a product truly worth a $3 million investment.

Eric Schultz and Andrew Wickerham are analysts for New America’s Health Policy Program and blog at The New Health Dialogue.

Submit a guest post and be heard on social media’s leading physician voice.

email

Comments are moderated before they are published. Please read the comment policy.

  • Anonymous

    An algorithm to predict hospitalization is a good idea.  I’m surprised it has already been done, if indeed it hasn’t.  Agree that geography is a key variable.  Any statisticians out there?  Please chime in.

Most Popular