Mayo Clinic ML Models Can Help Diagnose Hepatitis, Bacterial Infections

A study published in Mayo Clinic Proceedings shows that machine-learning (ML) models can assist healthcare workers with differentiating between alcohol-associated hepatitis and acute cholangitis using routine clinical variables.

The two conditions are very different, but they can present similarly in clinical settings such as the emergency department (ED) or the intensive care unit (ICU), the researchers noted. Alcohol-associated hepatitis is liver inflammation caused by drinking alcohol and can cause serious complications such as cirrhosis, kidney failure, or enlarged veins. Acute cholangitis is a bacterial infection usually associated with gallstones and results from inflammation of the bile duct system.

In both conditions, symptoms include fever, jaundice, right upper quadrant pain, and elevated liver enzymes, which can present a challenge to ED or ICU staff trying to diagnose and treat patients. To address this challenge, Mayo Clinic researchers developed ML models that can distinguish between the conditions using commonly available lab values.

To develop the models, researchers analyzed EHRs from 459 patients over 18 years admitted to Mayo Clinic in Rochester, Minnesota, and diagnosed with acute cholangitis or alcohol-associated hepatitis between Jan. 1, 2010, and Dec. 31, 2019. At the time of admission, 10 lab values were collected for each patient: white blood cell count, hemoglobin, mean corpuscular volume, platelet count, aspartate aminotransferase, alanine aminotransferase, alkaline phosphatase, total bilirubin, direct bilirubin, and albumin.

After records with incomplete data were removed, 260 patients with alcohol-associated hepatitis and 194 with acute cholangitis remained, whose records were used to train eight ML algorithms. The models were externally validated using data from a cohort of ICU patients at Beth Israel Deaconess Medical Center from 2001 to 2012.

During internal validation, the models achieved 93 percent accuracy in discriminating between the two conditions. During external validation, this rate fell to 90 percent.

To further evaluate the models, their performance was compared to 143 physicians who took an online survey based on the same 10 routine clinical variables to measure their performances. Overall, these clinicians achieved an accuracy of 79 percent, indicating that the ML models achieved significantly higher accuracy in both internal and external validation.

"The study highlights the potential for machine-learning algorithms to assist in clinical decision-making in cases of uncertainty," stated Joseph Ahn, MD, a third-year gastroenterology and hepatology fellow at Mayo Clinic in Rochester and first author of the study, in the press release. "There are many instances of gastroenterologists receiving consults for urgent endoscopic retrograde cholangiopancreatography in patients who initially deny a history of alcohol use but later turn out to have alcohol-associated hepatitis. In some situations, the inability to obtain a reliable history from patients with altered mental status or lack of access to imaging modalities in underserved areas may force providers to make the determination based on a limited amount of objective data."

The researchers also noted that ML algorithms such as these, if made easily accessible via an online calculator or smartphone app, could lead to improved diagnostic accuracy and reduced numbers of additional tests or inappropriate ordering of invasive procedures, which can delay correct diagnoses or subject patients to the risk of unnecessary complications.