Fatty Liver Screening Is a Trigger Problem, Not a Model Problem

Researchers quoted by Wired are pitching AI as a way to find fatty liver disease early, in a population Wired describes as roughly 30 percent of adults worldwide, or more than a billion people ^[1][2]. The interesting part of that pitch is not model performance. It is that the disease supplies no event that causes anyone to order a test.

The clinical picture is specific. Fat is negligible in a healthy liver, but in many adults and children it exceeds 5 or even 10 percent of the organ's weight, producing inflammation, cell damage, and the scarring known as fibrosis ^[3][4]. Left alone it can end in liver failure, and it has been linked to higher risk of cardiovascular disease and several cancers ^[5]. It typically develops without noticeable symptoms, so it is rarely caught early ^[6]. Even at the cirrhosis stage, three-quarters of people are diagnosed only once the condition is life-threatening ^[7], which leaves about one in four found before that point ^[8].

That is a detection funnel with no entry point. The tests themselves are not the constraint. The Fib-4 index scores advanced fibrosis risk from 0 to 6 using a patient's age, two liver enzyme levels, and blood-clotting ability ^[9]. Those inputs come from a liver blood test that in the US is often already run at an annual checkup ^[10], which means for many patients the score is arithmetic over data the chart already holds, with no new specimen required ^[11]. A second-line enhanced liver fibrosis test, measuring two scar-forming proteins and an enzyme that blocks scar clearance, is also available ^[12], and using both tests in patients with worrying liver fat has been shown to improve diagnosis of advanced fibrosis four-fold ^[13]. And yet these noninvasive assessments are rarely used, even in people at clearly elevated risk such as those with obesity or type 2 diabetes ^[14].

The reason offered is workload. Physicians facing growing administrative burden do not treat additional testing as sustainable ^[15]. Jonathan Dranoff of Yale puts the design requirement plainly: it has to run in the background, or be one button ^[16]. That is why Dranoff and Jeffrey Lazarus of the CUNY Graduate School of Public Health and Health Policy both foresee AI automating Fib-4 calculation from routine bloodwork so primary care can refer the right patients ^[17]. Lazarus frames the electronic health record as the substrate: AI can go back through massive numbers of hospital visits and lab reports to prioritize who is most at risk ^[18]. Note the verb. This is a role they anticipate, not a deployed system with published outcomes.

Accuracy is a secondary argument, and the numbers show why. An Osaka Metropolitan University model published last year read routine chest x-rays, which are aimed at lungs and heart but also capture part of the liver, and identified fatty liver disease with 82 percent accuracy ^[19][20]. That is roughly one in five classified wrong ^[21], which would be poor as a diagnosis and adequate as a prompt in a pathway where the current default is no test at all.

Screening only pays if something follows it. Early on, much of the damage is reversible through less alcohol, weight loss via diet and exercise, and more coffee ^[22]; at moderate to advanced scarring, semaglutide and resmetirom have been shown to be effective ^[23]. Lazarus argues the field's habit has been late-stage care and survival time rather than early detection ^[24].

Watch whether Fib-4 auto-calculation ships default-on inside EHR vendors rather than as an opt-in module, whether flagged patients actually reach hepatology, and whether anyone reports downstream diagnoses rather than alert volume.