AI models offer potential low-risk alternative to liver biopsy
Using machine learning, researchers at Beth Israel Deaconess Medical Center have developed a less-invasive alternative to liver biopsy—considered the gold standard for the diagnosis of non-alcoholic fatty liver disease.
The problem with liver biopsy is that it is a costly method that carries risks for the patient and is not always representative of the actual status of the entire liver.
“It would be impossible to do biopsies in the 80 million at-risk Americans, and even if we did, it would result in tens of thousands of subjects suffering complications and about 16,000 deaths each year from complications,” says Christos Mantzoros, MD, director of the Human Nutrition Unit at BIDMC and professor of Medicine at Harvard Medical School. “Finding easy-to-obtain, relatively inexpensive and reliable biomarkers which can be measured with less invasive techniques is an urgent, unmet need.”
Towards that end, BIDMC endocrinologists have generated novel predictive models—with the use of different machine learning methods—for the non-invasive diagnosis of non-alcoholic steatohepatitis and liver fibrosis based on measurements of lipids, glycans and biochemical parameters in peripheral blood.
The models may serve as low-risk, cost-effective alternative methods to liver biopsy for diagnosing and monitoring non-alcoholic fatty liver disease, according to results of a proof-of-concept study published in the journal Metabolism, Clinical and Experimental.
“We measured as many circulating molecules as reasonably possible, and then let machine learning and artificial intelligence pick the best sets of molecules that would most accurately predict outcomes,” adds senior author Mantzoros. “Although the number of subjects appears small, given conventional study designs, employing powerful and novel artificial intelligence models allowed us to derive accurate results, as high as 98 percent in some cases.”
According to the study’s authors, the models “should be further trained prospectively and validated in large independent cohorts” with more ethnically diverse participants, as well as include additional variables useful for enhancing the models’ predictive capabilities—such as patients’ genetic profile and clinical information including age and BMI,