A new machine learning study suggests that liver cancer risk may be predicted with a high degree of accuracy using routine clinical information such as age, medical history, and standard blood tests.

The findings are notable for two reasons. First, the model identified more true cases of hepatocellular carcinoma (HCC), the most common form of liver cancer in adults, than existing risk tools. Second, adding more complex data — including genomic information — did not appear to improve performance. That suggests widely available clinical data may be sufficient to build an effective early warning system.

The study was published in Cancer Discovery.

Why earlier detection matters

Liver cancer is the sixth leading cause of cancer death in the United States. HCC accounts for most adult liver cancer cases and is usually linked to chronic liver disease, including cirrhosis and viral hepatitis.

One of the major clinical challenges is that HCC often causes no symptoms in its early stages. As a result, many patients are diagnosed only after the cancer has advanced, when treatment options may be more limited.

Current screening strategies mainly target people already known to have chronic liver disease. But that approach does not capture everyone at risk. Researchers note that about 20% of HCC cases may occur in people without documented liver disease, meaning they may never qualify for routine surveillance under current guidance.

That gap has increased interest in tools that can identify high-risk individuals earlier and more broadly.

A machine learning approach to risk prediction

To explore whether artificial intelligence could improve early risk detection, researchers developed a machine learning model using data from the UK Biobank, a large health database containing information from more than 500,000 people.

Within that dataset, investigators identified 538 cases of HCC. Nearly 70% of those cases occurred in individuals who did not have a prior diagnosis of cirrhosis or chronic liver disease.

The research team trained the model on 80% of the dataset and then tested it on the remaining 20%. To see whether the tool would perform well in a different population, they also conducted external validation using the U.S.-based All of Us research program, which includes more than 400,000 participants and a more diverse study population. In that dataset, researchers identified 445 HCC cases.

Routine clinical data performed best

The model used a “random forest” method, a type of machine learning algorithm that combines multiple decision trees to improve prediction.

Researchers tested several versions of the model using different combinations of data. The best-performing version combined:

  • demographic information
  • electronic health record data
  • routine laboratory test results

This version, referred to as Model C, achieved an AUROC of 0.88, a level generally considered strong discrimination. In practical terms, this means the model was highly effective at distinguishing people who developed HCC from those who did not.

Importantly, more sophisticated inputs such as genomic data did not substantially improve results. That finding suggests a simpler and potentially more scalable approach may be possible.

Better performance than existing tools

The researchers also compared the new model with existing liver-related clinical tools, including FIB-4, APRI, NFS, and aMAP. These scores are often used to estimate liver fibrosis or cancer risk, especially in patients already known to have liver disease.

In this study, the machine learning model outperformed those tools overall. It identified more true HCC cases while also producing fewer false positives.

That balance matters. A tool that misses too many true cases may fail to detect people who need follow-up. A tool that produces too many false alarms may lead to unnecessary testing and anxiety.

To improve real-world usability, the researchers then simplified the model further. The final version relied on just 15 routinely collected clinical features and still performed better than the existing comparison tools.

Possible role in primary care

The study suggests the model could potentially be used as a prescreening tool in primary care settings. Rather than replacing specialist evaluation, it could help clinicians identify which patients may benefit from closer liver assessment or referral for screening.

This may be especially useful for patients who would otherwise be missed because they do not yet carry a diagnosis of chronic liver disease.

If validated in real-world practice, the approach could help shift some HCC diagnoses toward earlier stages, when curative treatment options are more likely to be available.

Important limitations

Although the results are promising, the study has limitations.

The analysis was retrospective, meaning it relied on existing data rather than testing the model prospectively in live clinical care. The number of participants with viral hepatitis — one of the major causes of HCC worldwide — was also relatively small.

In addition, although the model held up reasonably well in a more diverse U.S. population, most of the original training data came from white participants in the UK Biobank. That means broader validation is still needed before the tool can be considered ready for routine use across different healthcare systems.

What comes next

The next step is prospective, multi-center validation. Researchers want to determine whether the model can reliably identify patients who truly need additional liver-focused care when used in day-to-day clinical practice.

Because HCC is relatively uncommon, large health systems will likely be needed to test the model effectively over time.

For now, the study adds to growing evidence that AI may help improve early cancer detection — not necessarily through complex or expensive testing, but by making better use of information already available in routine care.