U doctors create algorithm to predict grave illness to boost care

Doctors are often fooled in making prognoses for severely ill patients in hospitals, so the University of Minnesota has created a predictive algorithm to try to take out the guesswork.

Studying medical records and histories of 60,000 patients admitted to Fairview hospitals, Dr. Nishant Sahni and colleagues found predictors that separated the patients who had longer life expectancies from those who were likely to die within one year. In a study published late last month, the doctors found they could predict within 10 percentage points a person's one-year death risk.

"We do this as physicians already — we risk-stratify these patients on the backs of our hands," said Sahni, a hospitalist and adjunct professor at the U's School of Medicine. "This is just a mathematical way of doing it. Humans are very bad at probability. The machine is just very good."

What exactly doctors and patients do with risk predictions is another matter. But right now, the absence of analytical information leaves critically ill patients uncertain about their futures, and doctors hedging on whether to talk to patients about their wishes near the end of life.

Even after a decade of heightened attention to palliative and end-of-life care programs, studies find that many patients die in hospitals when they would rather die in their homes. And money continues to be wasted on futile care measures that don't really extend patients' lives.

"We spend so much more money, and the outcomes are not that great," Sahni said. "I'm not trying to say it should go one way or the other (in terms of whether doctors and patients pursue aggressive care). Hopefully the model just spurs conversations."

Sahni's study in the Journal of General Internal Medicine showed more than two dozen lab tests that combined to make accurate mortality predictions, including pulse rate, body mass, and levels of nitrogen and platelets in blood.

Predictive models exist for everything from liver and breast cancer prognoses, to the likelihood of being readmitted to hospital care. But Sahni said this is the first study to predict death risk that isn't specific to any one disease, and to do so based on such a broad and reliable database of patients.

Hospitals aren't using the algorithm yet, though the university's office for technology commercialization is marketing it. Sahni said the study was a proof of concept, and that he is seeking grants and hospital partners nationally to test the algorithm on future patients.

Jeremy Olson • 612-673-7744