Background
Bilateral hepatocellular carcinoma refers to liver cancer involving both sides of the liver. In clinical practice, this often reflects multifocal disease that may arise through different biological pathways, making prognosis and treatment selection highly variable. Some patients can benefit from liver resection, while others are better served by a combined or non-surgical approach. The major challenge is identifying which patients are likely to do well after surgery and which patients have more aggressive disease despite technically successful resection.
This study used machine learning to explore whether patients with bilateral hepatocellular carcinoma could be separated into distinct clinical phenotypes based on preoperative and operative characteristics. The goal was not only to describe patterns of disease, but also to see whether those patterns predicted survival and recurrence more accurately than traditional single-factor assessment.
Study Design and Methods
Investigators reviewed a multi-institutional database of patients who underwent curative-intent hepatectomy for bilateral hepatocellular carcinoma between 2000 and 2023. Curative-intent hepatectomy means the operation was performed with the goal of removing all visible tumor. Patients were analyzed using k-means clustering, an unsupervised machine learning method that groups patients with similar features without pre-labeled categories.
The researchers then compared the resulting subgroups by tumor burden, operative approach, pathology, overall survival, and recurrence-free survival. Overall survival refers to the time from surgery to death from any cause. Recurrence-free survival refers to the time from surgery until cancer returns or the patient dies, whichever comes first. Survival outcomes were estimated with Kaplan-Meier methods and tested with multivariable Cox regression, which adjusts for multiple clinical factors at the same time.
Main Findings
A total of 347 patients were included. Machine learning identified two distinct phenotypes.
Cluster 1 included 95 patients and was marked by larger tumors, greater overall tumor burden, non-cirrhotic livers, and a higher likelihood of major hepatectomy. Major hepatectomy means a larger liver resection, often required when disease is extensive or centrally located.
Cluster 2 included 252 patients and was characterized by multiple smaller lesions. These patients were more often treated with minimally invasive surgery and with concomitant ablation, meaning heat-based tumor destruction performed at the same time as surgery.
The two clusters were not simply different in size; they also represented different biologic and clinical patterns. Cluster 1 had more aggressive pathologic features and substantially worse outcomes. Five-year overall survival was 34.9% in cluster 1 compared with 63.2% in cluster 2. Early recurrence was also more common in cluster 1, occurring in 53.7% versus 38.1% in cluster 2.
Importantly, cluster classification remained independently associated with prognosis even after accounting for other high-risk variables such as surgical margin status and microvascular invasion. Surgical margin status refers to whether the tumor was fully removed with a rim of normal tissue. Microvascular invasion means cancer cells are found in small blood vessels near the tumor, a known marker of aggressive disease. In the adjusted analysis, the cluster assignment predicted both overall survival and recurrence-free survival.
Clinical Interpretation
This study suggests that bilateral hepatocellular carcinoma is not a single disease entity. Instead, it appears to include at least two clinically meaningful phenotypes. One phenotype is dominated by larger, more extensive tumors and tends to require more complex surgery. The other phenotype involves multiple smaller lesions and seems more compatible with less invasive or combined local therapies.
From a surgical planning perspective, this distinction is important. A patient with low-burden multifocal disease may achieve favorable long-term survival after resection, especially when surgery is combined with ablation in selected lesions. By contrast, patients with dominant tumor plus satellite lesions or very high tumor burden may have a worse prognosis even after technically successful surgery and may benefit from multimodal treatment strategies, such as surgery combined with locoregional therapy or systemic therapy, depending on liver function and tumor biology.
The fact that phenotype remained predictive after adjustment for pathology underscores the value of preoperative risk stratification. In other words, the pattern of disease seen before surgery may already reflect important biologic differences that are not fully captured by pathology alone.
Why Machine Learning Matters
Traditional cancer studies often compare one factor at a time, such as tumor size, number of lesions, or vascular invasion. While useful, this approach may miss complex interactions among variables. Machine learning clustering can identify hidden patterns across multiple features and group patients into meaningful categories that better reflect real-world heterogeneity.
In this study, k-means clustering helped reveal two subtypes of bilateral hepatocellular carcinoma with distinct treatment patterns and outcomes. This type of analysis may eventually support precision surgery, where treatment is tailored not only to the presence of cancer but to the specific phenotype of the disease.
Implications for Practice
The findings may help clinicians refine surgical selection for patients with bilateral hepatocellular carcinoma. Preoperative variables such as tumor size, number of lesions, liver condition, and expected surgical extent could be used to estimate whether a patient belongs to a more favorable or unfavorable phenotype.
In practical terms, this may influence several decisions:
First, surgeons may choose between major resection, limited resection, minimally invasive techniques, and combined ablation strategies based on the likely phenotype.
Second, patients with more aggressive disease may be considered earlier for additional therapy after surgery or in combination with surgery.
Third, the model may help counsel patients more realistically about recurrence risk and long-term outcomes.
These findings do not mean that surgery should be avoided in higher-risk patients. Rather, they support more individualized care and highlight the need for multimodal planning in selected cases.
Limitations
As with any retrospective study, there are important limitations. The data were collected from multiple institutions over a long time span, and surgical techniques, imaging quality, and adjuvant treatment strategies may have changed over the years. In addition, clustering is an exploratory method; it identifies patterns but does not prove causation.
Another limitation is that machine learning models depend on the variables available in the database. Important biological markers, molecular data, and detailed treatment timing may not have been fully captured. Therefore, while the two clusters are clinically useful, they should be validated in independent cohorts before being used broadly in routine decision-making.
Conclusion
This study shows that bilateral hepatocellular carcinoma contains two clinically relevant phenotypes with different tumor burdens, operative strategies, and prognoses. Patients with low-burden multifocal disease had better survival, while those with dominant and satellite-type disease experienced worse outcomes. Preoperative phenotype-based stratification may improve surgical selection and help guide a more tailored, multimodal treatment approach for this challenging form of liver cancer.
