AI learns to spot fatty liver disease in routine blood tests and x-rays

Fatty liver disease is spreading quietly. Wired reports that more than a billion people worldwide now have livers accumulating fat: many adults and even children have livers where fat exceeds 5 percent, or in more severe cases 10 percent, of the organ's total weight, well above the negligible amount found in a healthy liver. That excess fat drives inflammation, cell damage and the scarring known as fibrosis, and the resulting condition, fatty liver disease, now affects about 30 percent of adults worldwide. Left untreated, it can progress to liver failure and has been linked to a higher risk of cardiovascular disease and several cancers. The trouble is that it usually causes no noticeable symptoms, so it is rarely caught early: even among people who reach cirrhosis or advanced scarring, three-quarters are not diagnosed until the disease has become life-threatening.

Catching it early matters because much of that early damage is highly reversible. In the initial stages, lifestyle changes, cutting back on alcohol, losing weight through diet and exercise, and even drinking more coffee, have all been shown to reverse scarring and inflammation, and for people with moderate to advanced scarring, the GLP-1 drug semaglutide and a medication called resmetirom have proven highly effective. Jeffrey Lazarus, a professor at the CUNY Graduate School of Public Health and Health Policy, puts it this way: 'The liver is a very versatile and interesting organ because it can regenerate, fibrosis can be reversed, and you can be completely healthy again. But traditionally, we've focused more on late-stage care and trying to see how long we can keep a patient alive rather than finding them early and preventing the condition from advancing.'

Simple, noninvasive ways to catch the disease early already exist, but one frustration for Lazarus and other specialists is that they are rarely used, even in people known to be at much higher risk, such as those with obesity and type 2 diabetes. The Fib-4 index, for example, scores a person's risk of advanced liver fibrosis on a scale of 0 to 6, based on their age, the levels of two liver enzymes, and their blood-clotting ability, drawing on a liver blood test that is often already part of an annual US checkup. Doctors also have access to a more accurate second-line blood test, which measures two proteins involved in forming scar tissue plus an enzyme that inhibits scar clearance; using both tests together in patients with worrying amounts of liver fat has been shown to improve diagnosis of advanced fibrosis four-fold. But for physicians already facing a growing workload and administrative burden, simply adding more testing to the routine is not seen as sustainable. 'You have to have something that can run in the background or it's easy to just hit a button and do it,' says Jonathan Dranoff, a professor of medicine at Yale University.

That gap is what researchers want AI to fill. Dranoff and Lazarus both expect AI to take data already collected during routine blood testing and automatically calculate Fib-4 scores, making it easier for primary care doctors to identify which patients to refer to a liver specialist. A second route runs through imaging that was never meant to check the liver at all: last year, scientists at Osaka Metropolitan University in Japan published a study showing that an AI model trained to analyze routine chest x-ray scans, taken mainly to examine the lungs and heart, could identify people with fatty liver disease with 82 percent accuracy, because those scans also pick up parts of the liver. Lazarus wants that kind of detection folded into every routine x-ray read where the liver happens to appear: 'The AI could pick up cases of excess liver fat, check for other risk factors such as if the person is overweight, has high cholesterol or type 2 diabetes, and then make a recommendation to the doctor,' he says. 'It could tell them this might not have been what you were looking for, but this is what was picked up, and you should refer to hepatology or endocrinology who can do further tests.'

AI is also being built to beat Fib-4 at its own game. Fib-4 is quick and low-cost, but it is less accurate in certain age groups, such as adolescents and seniors, and studies have raised concerns about false positives and unnecessary referrals unless it is combined with additional testing. The Danish health tech startup Evido has built an AI algorithm called LiverPRO that assesses a patient's risk of liver fibrosis from their age and nine routine blood-based biomarkers; now being commercialized in partnership with the pharmaceutical company Roche, it has been shown to outperform Fib-4 at predicting the risk of serious liver problems in more than 470,000 middle-aged people. A separate AI model called ALADDIN, another algorithm built on routine blood tests, was evaluated in a study published earlier this year by an international collective of hepatologists; it outperformed Fib-4 and other risk scores at identifying which patients would benefit most from resmetirom, without requiring an invasive liver biopsy first.

Paul Brennan, a specialty registrar in gastroenterology, hepatology, and internal medicine at the University of Dundee, is careful about how far these tools should go. 'These tools won't completely replace biopsies or imaging,' he says. 'But they could fix the bottlenecks in primary care where most fibrosis goes undetected. I expect them to be adopted as a smarter first pass, catching the moderate-risk patients that blunter tools may miss, and reducing unnecessary referrals to hepatologists.' For now, the use of AI in liver care has largely been confined to research rather than routine clinical practice, though Lazarus is optimistic that will start to change.

Part of that optimism rests on evidence beyond the tools themselves. Lazarus points to research carried out in Denmark that found informing people they have liver fibrosis makes them more likely to commit to diet and exercise changes. 'We're always looking to improve adherence to eating well and doing more physical activity,' he says. 'Telling someone that they might have or do have liver disease is one way to do that.' He also makes an economic case for finding patients sooner: 'I would love to say, let's go back through the electronic medical records across the various US health systems and find people before they have cirrhosis,' he says. 'Liver transplants are extraordinarily expensive in any country, but especially the US. So there's a lot of good humane and economic reasons to find people earlier on.'

Key facts

Fatty liver disease affects about 30 percent of adults worldwide and usually causes no symptoms; even among people who reach cirrhosis or advanced scarring, three-quarters are not diagnosed until the disease is life-threatening.
An AI model trained on routine chest x-rays, published last year by scientists at Osaka Metropolitan University in Japan, identified fatty liver disease with 82 percent accuracy, even though the scans were taken mainly to check the lungs and heart.
LiverPRO, an AI algorithm from the Danish startup Evido that scores risk from a patient's age and nine blood biomarkers, is now being commercialized with the pharmaceutical company Roche and has outperformed the standard Fib-4 test in more than 470,000 middle-aged people.
A separate AI model called ALADDIN, evaluated in a study published earlier this year by an international collective of hepatologists, outperformed Fib-4 at identifying which patients would benefit most from the drug resmetirom, without requiring an invasive liver biopsy.
Jonathan Dranoff and Jeffrey Lazarus want AI to automate the existing Fib-4 risk score from data already collected in routine blood tests, since adding new tests to an already stretched clinical workflow is not seen as sustainable.

Why it matters

Fatty liver disease is common, usually silent until it is advanced, and expensive to treat once it reaches that stage, which is exactly the profile of a problem that benefits from earlier detection. The article's case is that AI does not need new infrastructure to help: it can mine data that already exists, electronic health records, routine blood tests, even chest x-rays ordered for an unrelated reason, to flag people at risk before symptoms appear. That matters because much of the early damage from fatty liver disease is reversible through lifestyle changes, and newer drugs such as semaglutide and resmetirom are highly effective once scarring has progressed further; catching the disease earlier turns a currently expensive, late-stage problem into a treatable one.

Who it affects

Fatty liver disease affects about 30 percent of adults worldwide, including many adults and even children whose liver fat exceeds 5 percent, or in more severe cases 10 percent, of the organ's total weight; people with obesity or type 2 diabetes carry a much higher risk. Primary care physicians are the other pressure point: the article describes them as already stretched by workload and administrative burden, which is why automating existing tests rather than adding new ones is the approach researchers favor. Jeffrey Lazarus (CUNY Graduate School of Public Health and Health Policy), Jonathan Dranoff (Yale University) and Paul Brennan (University of Dundee) all appear as commentators on this trend; the article does not credit any of them with personally building Fib-4, LiverPRO or ALADDIN. The companies and researchers actually behind the named tools are the Danish startup Evido and its partner Roche for LiverPRO, scientists at Osaka Metropolitan University in Japan for the chest x-ray model, and an international collective of hepatologists for ALADDIN.

How to use it

Two of the approaches described are already usable in some form. LiverPRO, built by Evido from a patient's age and nine routine blood biomarkers, is now being commercialized in partnership with Roche. Fib-4 itself is already available today: it is computed from a liver blood test that is often already part of an annual US checkup, and Dranoff and Lazarus both want that calculation automated so physicians do not have to add a manual step to an already busy workflow. Combining Fib-4 with the second-line enhanced liver fibrosis blood test, without any AI involved, already improves diagnosis of advanced fibrosis four-fold in patients with worrying liver fat levels. The x-ray-based approach and ALADDIN remain further from bedside use: the article states plainly that AI's role in liver care so far has largely been confined to research. No price is given anywhere in the article for any of the tests, drugs or AI tools it describes: Fib-4, the enhanced liver fibrosis test, LiverPRO, ALADDIN, semaglutide or resmetirom.

How solid is it

The strongest evidence cited is LiverPRO's, tested against Fib-4 in more than 470,000 middle-aged people, and the Osaka Metropolitan University model's stated 82 percent accuracy at identifying fatty liver disease from routine chest x-rays. ALADDIN's evaluation, run by an international collective of hepatologists, is described only as outperforming Fib-4 and other risk scores, with no accuracy figure attached. The article names no journal, publisher or citation for any of the three studies, and gives only relative timing, 'last year' for the Osaka study and 'earlier this year' for ALADDIN, rather than a specific publication date. It also states no accuracy or sensitivity figures for Fib-4 or the enhanced liver fibrosis test on their own, only for the four-fold improvement that comes from combining the two.

Risks and caveats

Fib-4, the standard test most of these AI approaches aim to match or beat, is already known to be less accurate in certain age groups, such as adolescents and seniors, with studies raising concerns about false positives and unnecessary referrals unless it is combined with additional testing. Brennan is explicit that none of this replaces existing diagnostic tools: 'These tools won't completely replace biopsies or imaging,' he says, framing them instead as a smarter first pass that could catch moderate-risk patients and cut unnecessary hepatology referrals. The article gives no timeline for when AI might move from research into routine clinical use, stating only that such use has so far 'largely been confined to research,' and it does not say whether LiverPRO or ALADDIN have received any regulatory approval for clinical use.

“These tools won't completely replace biopsies or imaging. But they could fix the bottlenecks in primary care where most fibrosis goes undetected. I expect them to be adopted as a smarter first pass, catching the moderate-risk patients that blunter tools may miss, and reducing unnecessary referrals to hepatologists.”

— Paul Brennan, specialty registrar in gastroenterology, hepatology, and internal medicine, University of Dundee