- A machine learning model accurately predicted the risk of hepatocellular carcinoma (HCC) using routine clinical data.
- The model outperformed existing liver cancer risk tools by identifying more true cases while reducing false positives.
- The study suggests that adding complex data, such as genomics, did not improve performance, indicating that simple, widely available clinical data are sufficient for effective risk prediction.
- The tool could help clinicians detect at-risk individuals earlier, including those without diagnosed liver disease, potentially improving screening and patient outcomes if further validated.
It is not uncommon for people to receive a late-stage diagnosis of HCC. This is because it is usually asymptomatic in early stages. Current screening guidelines primarily focus on individuals with existing chronic liver disease.
However, roughly 20% of HCC cases may develop in those without any evidence of liver disease. Thus, these individuals are also at risk of a late diagnosis due to not meeting the criteria for surveillance.
Early diagnosis of HCC is essential, as many who receive a late diagnosis may not be suitable for current treatment options.
There is growing interest in the potential application of artificial intelligence (AI) for the early detection of HCC. Now, a new study, published in Cancer Discovery, suggests that a machine learning tool is capable of predicting HCC risk with high accuracy.
Although underlying liver disease is known as the most common risk factor for HCC, evidence highlights the role of other factors, such as being male, smoking, and heavy alcohol use. As multiple factors can influence HCC risk, identifying at-risk individuals has remained a challenge in clinical practice.
To address this, a research team led by Carolin Schneider, MD, an assistant professor of RWTH Aachen University, turned to machine learning, a form of AI that can analyze complex datasets and identify patterns across multiple variables simultaneously.
The researchers used data from the UK Biobank, which includes health information from more than 500,000 individuals. Among these participants, 538 cases of HCC were identified. Nearly 70% of these cases occurred in people without a prior diagnosis of cirrhosis or chronic liver disease.
The machine learning model was trained on 80% of the dataset, and performed an initial validation on the remaining 20%.
To test the model in a broader population, the team also conducted an external validation using the All of Us research program. This included data from more than 400,000 individuals in the U.S. and includes a more diverse participant pool. The registry included 445 cases of HCC.
Schneider told Medical News Today about the potential impact of this tool: “We hope that our pre-screening can be used in primary care to triage who should receive extra hepatological care.”
“By potentially identifying more people at risk earlier, we can develop pathways to refer them to screening or surveillance. Hopefully, this will help us detect HCC at an earlier stage, as earlier detection for HCC is strongly related to more curative treatment options.”
– Carolin Schneider, MD
The machine learning model used a “random forest” approach. This describes an algorithm that combines the output of multiple decision trees to generate predictions. The researchers tested models built from different types of clinical data.
The most effective version, referred to as Model C, combined patient demographics, electronic health records, and routine blood test results.
The performance of these models was assessed by calculating the area under the receiver operating characteristic (AUROC). This is a performance metric that describes the algorithm’s ability to distinguish between two groups. In this case, those in the validation cohort with HCC versus those without.
The algorithm achieved an AUROC score of 0.88, with 1 being a perfect score. This indicates that the model has a high accuracy in distinguishing between patients with and without HCC.
Notably, adding more complex data, such as genomics, did not significantly improve performance. This suggests that it may be possible to predict HCC risk using simple, readily available clinical data without the need for more expensive tests.
The researchers also compared their model with common clinical tools, including FIB-4, APRI, NFS, and the aMAP score. Healthcare professionals typically use these models to assess liver fibrosis or cancer risk in those with known liver disease.
The results suggest the machine learning model performed better overall, identifying more true cases of HCC, while reducing false positives.
“Current surveillance approaches are largely based on cirrhosis, but this misses HCC cases as chronic liver disease and especially cirrhosis is often under diagnosed,” Schneider told MNT.
“Our model introduces a pre-screening approach on routine data as basic demographic information, lifestyle and diagnoses as well as routine laboratory tests. This approach allowed us to identify individuals at risk of HCC with better precision/recall than currently used scores in our tested cohort,” she noted.
To help make Model C more practical for routine clinical use, the team further simplified it by reducing the number of clinical features it examines. The simplified version examines just 15 routinely collected clinical features and still outperformed the existing models.
“We brought our final score in a shape so that it is easily transferable to other health systems, the top 15 model consists only of routinely measured parameters and we provide to code to run it on local servers,” Schneider added.
These findings suggest that the model could help primary care physicians identify those who may otherwise be overlooked under current screening guidelines and refer them for liver cancer screening.
This could be significant for HCC, which is often aggressive but more treatable when caught early.
Although Model C was primarily trained on data from white participants from the UK Biobank, it maintained strong performance when tested in more ethnically diverse populations in the All of Us dataset. This suggests the approach could be broadly applicable across different demographic groups.
“Our results support potential transportability of our model, but obviously we want to test our model in as many health systems as possible to see on which factors good transportability depends and to perform regional calibration and validation,” Schneider said.
While the findings are promising, the authors note several limitations of the study. These include the retrospective design and the relatively low number of participants with viral hepatitis, one of the main causes of HCC.
When asked about future plans for testing this model, Schneider told MNT: “We need a prospective multi-center validation that shows that our score does identify the patients that need hepatological care.”
“HCC incidence is low, but roll out in large health systems will help us prospectively validate our pre-screening. We have therefore made the score and full pipeline openly available, with the explicit aim of enabling independent testing and external validation across many health systems,” she added.
Schneider concluded: “We hope that multiple clinical sites will trial the model and are happy to support!”
While further research is still necessary to validate Model C in additional populations and real-world clinical settings, the results highlight the growing potential of AI in healthcare, particularly in improving early detection strategies for conditions, such as liver cancer.
