At just 57, a Philadelphia man with a failing heart faced a tough decision: How did he want to die?
He told his doctors about his deep faith and love of family. He told them he wanted to spend as much time as he could at home, with his children, not in a hospital on life support. They shared his wishes with everyone on his medical team, and carefully recorded them in his medical records so anyone involved in his care would have no doubt about what to do.
There was something unique about this palliative-care consultation, however. It was sparked not by a sudden downward trend in his condition, nor by a pessimistic physician, nor by his own lost hope.
Instead, he was among the first Philadelphians singled out by a new kind of predictive tool, software that uses millions of test results from thousands of patients to look for signs that death may be imminent.
Palliative Connect was designed to identify patients at Penn Medicine who have a high risk of dying within six months so they have time to prepare.
"The goal is to make sure that all the patients who have a serious illness have a conversation about their goals and their wishes and their priorities, and that that gets documented in the record to guide their health care," said Nina O'Connor, director of palliative care at Penn.
But the data scientists behind Palliative Connect set out with a very different goal in mind: saving lives by predicting which patients were likely to develop one particular life-threatening condition.
All told, sepsis is the most costly condition treated in U.S. hospitals, affecting one in three patients who die while hospitalized. Sepsis occurs when the body overreacts to infection. If it progresses into septic shock, organs begin to shut down. The young and the old are hit hardest, but anyone can get this ghastly disease.
Yet when pediatric ER physician Fran Balamuth tells parents their child may be septic, they often don't know what she's talking about.
"Finding sepsis is one of the main challenges of being an emergency department physician," said Balamuth, who co-directs the Pediatric Sepsis Program at the Children's Hospital of Philadelphia. "There are tens of thousands of kids who come to the emergency room at CHOP with a fever — the vast majority of them don't have sepsis. That needle-in-a-haystack problem keeps many of us up at night."
>>READ MORE: Sepsis: A stealthy, sudden killer
When it comes to diagnosing the disease, every minute counts. Its earliest symptoms — fever, low blood pressure, a lack of alertness — resemble the effects of much more common and less dangerous conditions.
At CHOP, Balamuth helped develop an early alert system for sepsis designed to flag at-risk kids within minutes of their being brought to the emergency room.
Since 2011, Umscheid and a team of data scientists at Penn have been designing increasingly sophisticated algorithms to try to spot sepsis before it progresses.
Their first iteration — Early Warning System 1.0 — gave doctors a 2½ hour heads-up if a patient's vital signs were taking a turn for the worse. Yet sometimes that decline was due to sepsis, sometimes it was not.
Version 2.0 performed much better. "When the tool was launched, it was predicting sepsis on average a day and a half out," Umscheid said.
The benefit came from machine learning, a process by which computers train themselves to find subtle patterns in data.
Training a computer in this way requires a lot of examples — 162,212 patient records, to be exact. The new algorithm also analyzed advanced biomarkers such as the levels of creatinine and bilirubin in the blood, which can offer clues about underlying infection.
Umscheid compared the algorithms with routine tests clinicians run when diagnosing a patient, such as an echocardiogram that may reveal heart disease.
"But this is different," he said. "This isn't a clinician running an individual test, this is a health-care system having screening going on for every patient in the background."
The tool was turned off when Penn hospitals migrated to a new electronic health record system in 2017 and, despite its power, it remains off.
The problem wasn't so much the predictions. It was what to do with them.
Treating sepsis, often with fluid and antibiotics, is complicated. "If at the time of the alert they are not having signs of sepsis, giving them fluids can be dangerous — particularly if someone has heart failure or kidney problems," Umscheid said. "On the same side, giving someone antibiotics without knowing what you're treating — it could be all harm and no benefit."
This particular way of computerizing medicine had hit a snag.
"The hype is all over the place," Umscheid said, riffing on the role of machine learning in medicine. "For these predictions to be of value to clinicians and patients, they have to be predicting things that people care about — that people need help with — and doing so at a time when people can act on that to change the course of care."
Michael Draugelis, chief data scientist at Penn Medicine, agreed. "There are all kinds of correlations in data that appear interesting to the data scientists but are completely meaningless clinically," he said.
Now, said Draugelis, "we're looking at this from the human side."
The data scientists partnered with Penn's palliative care program to develop Palliative Connect.
True, nobody needs a computer to tell them they should think about — and act on — such an important issue while they are well enough to do so. But only about a third of U.S. adults complete any type of advance directive for end-of-life care. A pilot study conducted last winter suggested Palliative Connect could help improve that.
The tool has started rolling out to departments at both HUP and Pennsylvania Hospital, including in the cardiac unit caring for the 57-year-old man in heart failure. O'Connor couldn't divulge more about him, but did say that Palliative Connect is being better received than some might have feared.
"We actually called patients after they were discharged from the hospital to find out their perception and to make sure that we weren't causing any harm or alarming patients and families," O'Connor said.