Analytics Take an Unusual Turn at PeraHealth

Posted on August 17, 2017 I Written By

Andy Oram is an editor at O’Reilly Media, a highly respected book publisher and technology information provider. An employee of the company since 1992, Andy currently specializes in open source, software engineering, and health IT, but his editorial output has ranged from a legal guide covering intellectual property to a graphic novel about teenage hackers. His articles have appeared often on EMR & EHR and other blogs in the health IT space.

Andy also writes often for O’Reilly’s Radar site (http://oreilly.com/) and other publications on policy issues related to the Internet and on trends affecting technical innovation and its effects on society. Print publications where his work has appeared include The Economist, Communications of the ACM, Copyright World, the Journal of Information Technology & Politics, Vanguardia Dossier, and Internet Law and Business. Conferences where he has presented talks include O’Reilly’s Open Source Convention, FISL (Brazil), FOSDEM, and DebConf.

Data scientists in all fields have learned to take data from unusual places. You’d think that monitoring people in a hospital for changes in their conditions would be easier than other data-driven tasks, such as tracking planets in far-off solar systems, but in all cases some creativity is needed. That’s what PeraHealth, a surveillance system for hospital patients, found out while developing alerts for clinicians.

It’s remarkably hard to identify at-risk patients in hospitals, even with so many machines and staff busy monitoring them. For instance, a nurse on each shift may note in the patient’s record that certain vital signs are within normal range, and no one might notice that the vital signs are gradually trending worse and worse–until a crisis occurs.

PeraHealth identifies at-risk patients through analytics and dashboards that doctors and nurses can pull up. They can see trends over a period of several shifts, and quickly see which patients in the ward are the most at risk. PeraHealth is a tool for both clinical surveillance and communication.

Michael Rothman, co-founder and Chief Science Officer, personally learned the dangers of insufficient monitoring in 2003 when a low-risk operation on his mother led to complications and her unfortunate death. Rothman and his brother decided to make something positive from the tragedy. They got permission from the hospital to work there for three weeks, applying Michael’s background in math and data analysis (he has worked in the AI department of IBM’s Watson research labs, among other places) and his brother’s background in data visualization. Their goal, arguably naive: to find a single number that summarizes patient risk, and expose that information in a usable way to clinicians.

Starting with 70 patients from the cardiac unit, they built a statistical model that they tested repeatedly with 1,200 patients, 6,000 patients, and finally 25,000 patients. At first they hoped to identify extra data that the nurse could enter into the record, but the chief nurse laid down, in no uncertain terms, that the staff was already too busy and that collecting more data was out of the question. It came time to get creative with data that was already being collected and stored.

The unexpected finding was that vital signs were not a reliable basis for assessing a patient’s trends. Even though they’re “hard” (supposedly objective) data, they bounce around too much.

Instead of relying on just vital signs, PeraHealth also pulls in nursing assessments–an often under-utilized source of information. On each shift, a nurse records information on a dozen different physical systems as well as essential facts such as whether a patient stopping eating or was having trouble walking. It turns out that this sort of information reliably indicates whether there’s a problem. Many of the assessments are simple, yes/no questions.

Rothman analyzed hospital data to find variables that predicted risk. For instance, he compared the heart rates of 25,000 patients before they left the hospital and checked who lived for a year longer. The results formed a U-shaped curve, showing that heart rates above a certain level or below a certain level predicted a bad outcome. It turns out that this meaure works equally well within the hospital, helping to predict admission to the ICU, readmission to the ICU, and readmission after discharge.

The PeraHealth team integrated their tool with the hospital’s EHR and started producing graphs for the clinicians in 2007. Now they can point to more than 25 peer-reviewed articles endorsing their approach, some studies comparing before-and-after outcomes, and others comparing different parts of the hospital with some using PeraHealth and others not using it. The service is now integrated with major EHR vendors.

PeraHealth achieved Rothman’s goal of producing a single meaningful score to rate patient risk. Each new piece of data that goes into the EHR triggers a real-time recalculation of the score and a new dot on a graph presented to the nurses. In order to save the nurses from signing into the EHR, PeraHealth put a dashboard on the nurse’s kiosk with all the patients’ graphs. Color-coding denotes which patients are sickest. PeraHealth also shows which patients to attend to first. In case no one looks at the screen, at some hospitals the system sends out text alerts to doctors about the most concerned patients.

PeraHealth is now expanding. In an experiment, they did phone interviews with people in a senior residential facility, and identified many of those who were deteriorating. So the basic techniques may be widely applicable to data-driven clinical decision support. But without analytics, one never knows which data is most useful.