AI Tool Identifies Undiagnosed Early-Stage MASLD

SAN DIEGO — An artificial intelligence (AI)–driven algorithm may be able to accurately detect early-stage metabolic dysfunction–associated steatotic liver disease (MASLD) based on imaging findings and other criteria in patient electronic medical records, according to new research.

Among the patients identified by the algorithm as meeting the criteria for MASLD, only a small percentage had an MASLD-associated diagnostic code.

“A significant portion of patients who meet criteria for MASLD go undiagnosed, which can lead to delays in care and progression to advanced liver disease,” said lead author Ariana Stuart, MD, an internal medicine resident at the University of Washington, Seattle, who presented the findings (abstract 2360) at The Liver Meeting 2024: American Association for the Study of Liver Diseases (AASLD).

“However, people shouldn’t interpret our findings as a lack of primary care training or management,” she said. “Instead, this study indicates that AI can complement physician workflow and address the limitations of traditional clinical practice.”

Developing an MASLD Algorithm

Typically, the identification of MASLD has relied on clinician recognition and descriptions in chart notes, Stuart said. Early-stage disease often goes unnoticed, particularly if patients remain asymptomatic, until cirrhosis develops.

To address this, Stuart and colleagues created a machine learning, natural language processing AI algorithm on the basis of MASLD criteria from AASLD: Hepatic steatosis on imaging and at least one metabolic factor (elevated body mass index, hypertension, prediabetes or diabetes, or dyslipidemia). The model was validated by two physicians, who manually reviewed monthly cohorts generated by the algorithm.

Between December 2023 and May 2024, the researchers used the algorithm to analyze an MASLD cohort from medical centers in the Seattle area. The mean age was 51 years, 44% were women, and 68% were White. Those with alcohol-associated liver disease, metastatic malignancy, and autoimmune, genetic, and infectious causes of liver disease were excluded.

The algorithm identified 957 patients with imaging that matched MASLD criteria.

Among those, 137 patients (17%) identified by the algorithm had an MASLD-associated diagnostic code. For these patients, the mean time from initial imaging with steatosis to diagnosis was 33 days, according to patient records.

An additional 26 patients received an MASLD diagnosis during the study period, with a mean time to diagnosis of 56.2 days.

In terms of patient management, 245 patients (26%) had contact with a gastroenterologist or hepatologist based on documentation of a letter, phone call, or office visit. In addition, 546 patients (57%) were screened for hepatitis C.

After adjusting for an over-inclusion error rate of 12.8% and an overdiagnosis rate of 0.02%, the research team found 697 patients (83%) lacked a relevant diagnosis. After multiple iterations, the algorithm achieved an accuracy of about 88%, Stuart said.

Considering Future AI Use

Stuart and colleagues are now testing the algorithm in larger groups and across longer periods.

After that, they intend to implement a quality improvement program to increase awareness for clinicians and primary care providers, as well as train users on how to interpret and move forward with findings of hepatic steatosis in patient records.

For instance, future AI models could flag patients for additional testing, improve chart review, and aid in research efforts around cardiometabolic comorbidities associated with MASLD, she said.

Looking ahead, AI tools such as these represent what’s possible for advancements in research, patient care, and clinical workflows, said Ashley Spann, MD, assistant professor and transplant hepatologist at Vanderbilt University, Nashville, Tennessee, and director of clinical research informatics for Vanderbilt’s Gastroenterology Division.

“AI, in my view, is actually augmented intelligence,” she added. “We need to think about the people and processes involved.”

Spann, who spoke to Medscape Medical News about the use of AI tools in medicine in general, stressed the need for transparency in AI use, careful validation of input-output data, frameworks for machine learning models in medicine, and standardization across institutions.

“What we ultimately need is an infrastructure that supports the simultaneous deployment and evaluation of these models,” she said. “We all need to be on the same page and make sure our models work in multiple settings and make adjustments based on algorithmvigilance afterwards.”

Stuart reported no relevant disclosures. Spann serves on Epic’s hepatology steering board, which has focused on how to use AI tools in electronic medical records.

Carolyn Crist is a health and medical journalist who reports on the latest studies for Medscape Medical News, MDedge, and WebMD.