Fatty liver disease affects roughly 30 percent of adults worldwide, yet it often develops without obvious symptoms. By the time advanced scarring is discovered, treatment becomes more difficult and risks can include liver failure, cardiovascular disease and cancer. Researchers now see an opportunity for AI to identify warning signs hidden in health data that already exists.
The strongest promise is not a futuristic diagnosis made by a single model. It is a practical screening layer that calculates risk from routine blood work, searches electronic records or notices liver abnormalities in images captured for another purpose. Used carefully, those systems could direct more people toward specialist assessment while the damage may still be reversible.
Existing tests are useful but inconsistently applied
Early-stage liver inflammation and fibrosis can improve through changes including weight loss, exercise and reduced alcohol consumption. Treatments such as semaglutide and resmetirom may also help some people with more advanced scarring. The central problem is finding the condition early enough to act.
Clinicians already have non-invasive tools. The Fib-4 index estimates the risk of advanced fibrosis using age, two liver-enzyme measurements and blood-clotting ability. A second-line enhanced liver fibrosis test measures proteins and an enzyme associated with scar formation and clearance. Using both tests has been reported to improve detection of advanced fibrosis fourfold.
Yet adding manual calculations and testing to overloaded primary-care workflows is difficult. Specialists interviewed by WIRED suggest that software could calculate Fib-4 automatically from routine results and flag the patients most likely to need further examination. That would make AI less a replacement for clinical judgment than a background triage mechanism.
The limitation is that simple scores perform unevenly. Fib-4 can be less accurate for adolescents and older adults, while false positives may generate unnecessary referrals unless the result is combined with other evidence.
Models are searching blood records and incidental images
Electronic health records contain years of visits and laboratory results that could help identify people whose risk has gone unnoticed. Jeffrey Lazarus of the CUNY Graduate School of Public Health and Health Policy argues that AI could retrospectively prioritise likely cases across large hospital populations.
Routine images offer another route. Researchers at Osaka Metropolitan University trained a model to detect fatty liver disease from chest X-rays, which capture part of the liver even when clinicians are primarily examining the lungs and heart. Their study reported 82 percent accuracy. That figure is encouraging, but it does not by itself establish performance across different hospitals, equipment or patient populations.
Blood-based models are also advancing. Danish health technology company Evido developed LiverPRO, which uses age and nine routine biomarkers to assess fibrosis risk. The system, now being commercialised with Roche, reportedly outperformed Fib-4 in data from more than 470,000 middle-aged people.
Another model, ALADDIN, was evaluated by an international hepatology group as a way to identify patients most likely to benefit from resmetirom. Researchers reported better performance than Fib-4 and other risk scores. These systems remain decision-support candidates; they do not eliminate the need for imaging, biopsy or specialist review.
Screening value depends on what happens after a flag
An automated alert is useful only if health systems can confirm the result and offer appropriate care. A model that finds more moderate-risk patients could reduce missed disease, but excessive false positives could move the bottleneck from detection to hepatology clinics.
Deployment therefore requires thresholds tailored to the population, clear referral pathways and monitoring for age or demographic groups where performance changes. Clinicians also need to understand which evidence produced a recommendation rather than receiving an unexplained warning.
Much of the reported AI work in liver care remains research. Its economic appeal is nevertheless substantial: earlier intervention may avoid costly treatment for cirrhosis and transplantation, while giving patients a better chance to reverse damage.
The near-term opportunity lies in using information that health systems already collect but do not consistently connect. If validated in real workflows, AI could make liver screening quieter and more systematic—calculating in the background, identifying overlooked risk and leaving diagnosis and treatment decisions with clinicians who can examine the full patient.