Home Analysis Predicting in-hospital mortality

Predicting in-hospital mortality

A hospitalist develops a tool to identify high-risk transfers

March 2023

FOR THE PAST SEVEN YEARS, Neetu Mahendraker, MD, lead academic hospitalist for the Indiana University Health System in Indianapolis, has worked to develop a mortality prediction tool that hospitalists can use to identify patients at high risk of dying in the hospital in the next 30 days.

But several factors slowed that process down. For one, Dr. Mahendraker had to postpone (for two years) a pilot study of the predictive tool she validated during the pandemic. And like many hospitalist researchers, while she has sustained support from leadership and mentors, she has had to fit both her research and quality improvement projects—and she has several— into her full-time clinical practice.

Now, however, the data are back from the pilot program she finally launched, and Dr. Mahendraker plans to roll out the tool she developed to screen more patients in her academic center. What began as something she felt she needed in her own practice may eventually become a standardized score in EHRs.

“We want patients to benefit from hospice and palliative care much earlier.”


Neetu Mahendraker, MD
Indiana University Health System

“I’m very excited,” says Dr. Mahendraker. “This is a novel model that can facilitate earlier goals-of-care discussions with patients and families.”

Screening transfer patients
Dr. Mahendraker started thinking about what she came to call the inpatient mortality predictive tool in 2015. At that time, even with 10 years of hospitalist experience, “I realized it was still very difficult for me to guess if the death of even severely ill patients was impending.” She figures that she could predict which patients were at high risk of 30-day in-hospital mortality 30% or 40% of the time.

“Clinical judgment is always helpful,” says Dr. Mahendraker. “But I knew a predictive tool could supplement clinical judgment, helping front-line doctors feel more comfortable making such decisions.”

The challenge, she adds, is getting these patients to talk about goals of care sooner rather than later. Like many physicians, she says that she often backed off from those discussions because she didn’t have enough data or because different treating teams had conflicting opinions. “Too often, even patients at very high risk don’t receive hospice or palliative care consults until they’re almost dying.”

Dr. Mahendraker knew that all the factors in any predictive model, which would be embedded in the EHR, would have to be available electronically within 24 hours of admission. She also knew the first screening target would be transfer patients.

That’s because transfer patients in a tertiary care center like hers have the highest mortality rate of any admissions, she points out. And “their families come with expectations that their loved one will be saved, no matter what.” That puts a lot of burden on tertiary care centers to help patients and families gain perspective.

“Often, we can provide only limited resources or the patient has only limited treatment options.”

Improving the odds
Dr. Mahendraker found proven mortality-prediction models for specific specialties, including cardiology and nephrology. “But there was nothing for general front-line providers.”

She started with 30 variables from validated models and guidelines, which she and her research colleagues winnowed down to 20. Among those 20, she believes patients’ Charlson Comorbidity Index score is particularly important. When she and her team validated the mortality tool in 2016 and 2017, they found a positive predictive value of 68%.

“That raised my hopes,” she points out. “If I can predict only 40% of the time that a patient will die within 30 days, a tool that raises that ability to close to 70% is a great addition.”

Results of the validation study were published in the Journal of General Internal Medicine in January 2021. A “threshold score of -2.19 resulted in a sensitivity of 75.00% and specificity of 85.71%,” she and her colleagues wrote. Above that threshold, 34.4% of patients died within 30 days—but below that threshold, only 2.8% of them did.

Encouraging outcomes
The pilot program to test the tool was launched in August 2021 and finished in January 2022. Dr. Mahendraker and her team identified 84 transfer patients at high risk for 30-day inpatient mortality using the predictive tool as a screen. Among them, 42 served as controls while the other 42 were an intervention group.

Her results: Hospitalists initiated goals-of-care discussions in 21% of the patients in the intervention group vs. only 9% in the control group. Moreover, such discussions were completed within 72 hours of admission for 16% of patients in the intervention group compared to only 4% of controls.

More patients in the intervention cohort were admitted to comfort care than among controls. And “importantly, more patients in the intervention changed their code status,” says Dr. Mahendraker: 21% vs. 7% among controls.

“These outcomes are all very encouraging,” she notes. “We want patients to benefit from hospice and palliative care much earlier and to help them understand their disease process sooner.”

What drives buy-in?
While Dr. Mahendraker prepares her pilot data for publication, she’s pulling together another set of data from that program: a qualitative analysis on the hospitalists who volunteered for the pilot to compare with those who did not.

“We want to understand what drives buy-in from the primary doctors using this scoring tool,” she says. And while the physicians treating 21% of the patients in the intervention group initiated end-of-life conversations, Dr. Mahendraker interviewed all the hospitalists who participated in the pilot to find out why some didn’t start such discussions.

“The hospitalists all agreed these patients were at risk for mortality in the hospital,” she explains. “But some held off on serious illness communications because they were waiting for pending data.” That led to delays in the timing of such discussions. “Different physicians have different approaches to end-of-life care conversations,” Dr. Mahendraker says.

“While some initiate those earlier in the care, others want to start discussions later.” Drawing data about how much trust physicians put in the model “will help make it a more robust tool.”

Her next step is to roll the screening tool out to a larger group of transfer patients, with the eventual goal of screening all admissions, not just transfers.

Ultimately, she thinks the tool’s use could drive more accurate documentation and capture illness complexity. Dr. Mahendraker also believes it may eventually allow clinicians in her facility to rethink some incoming transfers who don’t have definitive treatment options available. That may allow doctors to offer patients who are too sick to benefit from tertiary care tele-consults in palliative and other specialty care instead.

Phyllis Maguire is Executive Editor of Today’s Hospitalist.

Published in the March/April 2023 issue of Today’s Hospitalist

Notify of
Inline Feedbacks
View all comments