Machine learning strategies outperformed traditional methods of predicting inpatient cirrhosis mortality, enabling easier identification of at-risk patients who need extra attention at the hospital, according to newly published research.
“It’s very easy to be a Monday morning quarterback and say, ‘This patient went to the ICU while they were in the hospital. This patient had dialysis while they were in the hospital.’ Literally everyone would know that patient is not going to do well. What we want is for clinicians to know on the day of admission … what is the likelihood that this patient is going to get better?” said Jasmohan S. Bajaj, MD, MS, a professor of medicine at Virginia Commonwealth University, in Richmond, and the study’s lead investigator.
The new machine learning approach was developed using the global CLEARED Consortium, which includes a cohort of 7,239 inpatients with cirrhosis (64% men; mean age, 56±13 years; median Model of End-Stage Liver Disease-Na score, 25). Participants were from 115 health centers in every populated continent of the world, in countries of varying income levels. A total of 808 patients (11.1%) died while in the hospital.
The investigators evaluated 69 variables including clinical variables such as history of hepatic encephalopathy or an acute kidney injury when arriving at the hospital, as well as hemoglobin, white blood cell and platelet levels. They also tracked whether people had been hospitalized for any reason in the six months prior to admission for cirrhosis and whether they had an infection in those six months, among other pre-admission variables. All variables were input into different predictive models that were then compared (Gastroenterology 2025 Jul 23. doi:10.1053/j.gastro.2025.07.015). The investigators explored which model most effectively predicted inpatient death from cirrhosis.
Machine learning outperformed the traditional approach of logistic regression at predicting inpatient cirrhosis mortality. The area under the curve for machine learning was 0.815 (95% CI, 0.785-0.844), compared with 0.773 (95% CI, 0.738-0.807) for logistic regression. This represented a 4.2-point advantage (P<0.001) for machine learning, which was a better predictor regardless of where someone lived.
“The machine leaning model is very easy to use and is freely available,” Dr. Bajaj said.
The researchers validated the model in 28,670 U.S. veterans with cirrhosis, of whom 4% (n=1,158) died in the hospital. In this cohort, machine learning again outperformed other predictive models, even when using just 15 variables rather than the 69 variables used in the CLEARED cohort. Those 15 variables are included in a freely available tool (silveys.shinyapps.io/app_cleared/).
Dr. Bajaj noted that the mortality rates in both cohorts were lower than those seen in other studies of patients with cirrhosis, but he added that the CLEARED data more accurately reflect real-world clinical experience.
“Prior studies have very selected, very high-mortality patients with cirrhosis,” Dr. Bajaj explained. But if the machine learning model identifies someone at lower risk who nonetheless could still die in the hospital, clinicians can use that information to prioritize care for those most at risk. With this approach, he said, “there is a hope that you can do something more,” Dr. Bajaj said.
Room for Improvement
“[The researchers] identified some common factors in the various countries and income groups that lead to higher inpatient mortality from cirrhosis,” such as acute kidney injury and hepatic encephalopathy, said Mark Russo, MD, MPH, a transplant hepatologist at Advocate Health in Charlotte, N.C.
“That’s pretty unique to this study, to identify those common factors in inpatient mortality across many countries,” added Dr. Russo, who was not involved in the study. He called the research a “tremendous effort,” given the number of patients evaluated by machine learning and the breadth of countries included.
“This data suggest maybe this patient requires more frequent monitoring, or we should move them into intensive care because they’re really at high risk for dying,” Dr. Russo said. In the most concerning cases, he noted, the model might inform conversations with patients and their loved ones about transitioning to palliative or hospice care.
With additional data not collected in this study, such as what proportion of patients died within seven days of admission and what they died from, clinical management could be even more precise, Dr. Russo added. “I think those are the next steps,” he said.
For example, he explained, data could indicate that people with cirrhosis die of sepsis in the hospital more than expected. If that is the case, clinicians could order more blood cultures and monitor for signs of sepsis more intensively.
These two variables—death within seven days and cause of death—would be applicable to all the countries in the study, Dr. Russo said, adding that it also would be helpful to know which patients were saved by a liver transplant in countries where such transplants are available.
—Marcus A. Banks
Drs. Bajaj and Russo reported no relevant financial disclosures.
This article is from the October 2025 print issue.