Predicting who might die while in the hospital is one of the hardest challenges for doctors caring for people with serious liver disease.
A Virginia Commonwealth University School of Medicine gastroenterologist, working with an international team known as the CLEARED Consortium, reported Wednesday that they developed an AI model that can help save lives by providing better prediction of which hospitalized liver patients are at greatest risk of dying. The team reported its work in the journal Gastroenterology .
Jasmohan Bajaj, M.D., who is also with the VCU Stravitz-Sanyal Institute for Liver Disease and Metabolic Health and the Richmond Veterans Administration Medical Center, said that if physicians implement the artificial intelligence model, doctors can act sooner, guide tough conversations with families and target their care for patients with cirrhosis. Bajaj is the corresponding author on the consortium's study, which involved thousands of patients worldwide, including more than 29,000 U.S. military veterans.
"This means a doctor can have more confidence about which patients need the most urgent care, which ones might need hospice discussions with family members, who could need transfer to better-equipped hospitals and which patients are likely to recover," Bajaj said. "Medically and nonmedically, we can better approach the patient if we have a better handle on the patient's condition."
Cirrhosis happens when the liver is so badly damaged by alcohol, hepatitis or excess fat that it can no longer work properly. Patients often face multiple hospital stays and dangerous complications, including severe infections, kidney failure and confusion known as hepatic encephalopathy. Once hospitalized, they face a high risk of death, yet predicting which patients are in greatest danger is difficult.
To see if AI could help physicians assess their hospitalized patients with cirrhosis, the team initially turned to a prospectively collected consortium database of detailed health information on more than 7,000 patients with cirrhosis. These patients were treated at 121 hospitals across six continents, from the United States to Asia and Africa. Researchers recorded why the patients were admitted, what complications they had, what treatments they received and whether they survived the hospital stay.
Using advanced machine learning tools, the team tested how well four different models could predict which patients would die in the hospital. One of the tools was a traditional statistical method, and the other three used newer machine learning approaches.
The team found that the Random Forest analysis model worked best and outperformed an older statistical method by picking up hidden warning signs. The Random Forest model had an accuracy score of 0.815, higher than the traditional logistic regression method, which scored 0.773.
Even when the team simplified the model down to the 15 most powerful risk factors, making it very accessible for use around the globe, it still performed better than the older method. The most critical factors were whether the cirrhosis patient was admitted for kidney failure, had serious brain complications or experienced infections, which raise the risk of dying in the hospital.
Once a patient is identified as at high-risk, Bajaj said, health providers can intensify care and management of complications to prevent a patient's condition from worsening. If a patient is on the borderline for liver transplant, they could be considered for a liver transplant sooner rather than later.
"High-risk patients could be shifted to another hospital for better treatment," Bajaj said. "And the last thing is, if it's likely they're on the path toward decline and possibly palliative care, those decisions can be made earlier, when the patient is still awake and alert and can participate in making them.
"On the other hand," Bajaj continued, "if the patient is low-risk, doctors can feel more confident focusing on recovery and discharge planning."
The team then tested the same model in about 29,000 U.S. military veterans with cirrhosis who were treated in VA hospitals. Even among those patients, who were older and mostly men, the AI model outperformed standard scoring systems, giving doctors a clearer idea of which patients were likely to survive or not.
The consortium is sharing the tool with hospitals that care for patients with advanced liver disease, both in the U.S. and abroad. The research team designed the tool for ease of use; physicians only need to enter the 15 most important patient details to get a reliable risk estimate.
"Better prediction means better planning," Bajaj said. "When we know who is most at risk, we can target treatment, talk to families early and focus our resources where they matter most."
Other VCU and Richmond VA researchers who participated in the research included Somaya Albhaisi, M.D.; Brian Bush; Nilang Patel, M.D.; Jawaid Shaw, M.D.; Scott Silvey; and Leroy Thacker, Ph.D.
Partial funding came from the VA. Other funding came from the National Center for Advancing Translational Sciences of the National Institutes of Health, under Award Number UM1TR004360.
