P4574 - Clustering Hepatocellular Carcinoma Patients Using Unsupervised Machine Learning and Its Effect on Survival: A Population-Based Study

Mohamed Mortagy, MBBCh¹, Umed Nadir, MBBS¹, Muhammad Al-Dulaimy, MBBS¹, Aya Abdelhameed, MBBCh², Marawan Elmassry, MBBCh³, Amr Elgehiny, MBBCh⁴, Ahmed Elsandouby, MBBCh^1
1Hampshire Hospitals NHS Foundation Trust, Winchester, England, United Kingdom; ²Alexandria University, Alexandria, Al Iskandariyah, Egypt; ³Texas Tech University Health Sciences Center, Lubbock, TX; ⁴MD Anderson Cancer Center, Houston, TX

Introduction: Hepatocellular carcinoma (HCC) has an annual incidence of 9.5 per 100,000 person-years in the USA. Clustering is an unsupervised machine learning technique that helps subgroup patients based on patterns and differences observed among them.

Methods: A total of 90,080 adult patients diagnosed with HCC from 2000 to 2021 were extracted from the U.S. Surveillance, Epidemiology, and End Results (SEER) database after excluding patients with missing data. These patients were clustered into three groups (0, 1, 2) using unsupervised machine learning with K-means clustering. Clustering effectiveness was evaluated using silhouette (S), Davies-Bouldin (DB), and Calinski-Harabasz (CH) scores. Principal Component Analysis (PCA) and T-distributed Stochastic Neighbor Embedding (T-SNE) were used for cluster visualization. SHapley Additive exPlanations (SHAP) identified the most influential clinical features for clustering. Sixty-month survival was estimated for each cluster using Cox regression and Kaplan-Meier (KM) statistics. The impact of chemotherapy, radiotherapy, and surgery on each cluster survival was visualized with Kaplan-Meier curves. A decision tree was generated using supervised machine learning to classify patients into one of the three clusters.

Results: The cohort characteristics and details for each cluster are shown in Table 1. T-SNE and PCA visualizations showed good clustering . SHAP analysis indicated that the four most important variables for clustering were tumor size, age, stage, and sex in order. Survival analysis revealed statistically significant differences among the clusters based on Cox regression, log-rank test, and KM statistics (Figure 1). The highest 60-month survival rate was observed in Cluster 0 (31.5%), followed by Cluster 2 (17.5%), and Cluster 1 (5.71%). The decision tree achieved an overall accuracy of 92% in clustering patients. Patients receiving surgery, chemotherapy, or radiotherapy demonstrated improved survival across all clusters. This study is limited by some incomplete and potentially inaccurate variables in the SEER database and its retrospective nature.

Discussion: Unsupervised machine learning could be useful for clustering HCC patients. This could help identify new subgroups or interactions between variables to be able to predict survival outcomes, and assess treatment effects within different patient clusters.



Disclosures:


Mohamed Mortagy, MBBCh¹, Umed Nadir, MBBS¹, Muhammad Al-Dulaimy, MBBS¹, Aya Abdelhameed, MBBCh², Marawan Elmassry, MBBCh³, Amr Elgehiny, MBBCh⁴, Ahmed Elsandouby, MBBCh¹. P4574 - Clustering Hepatocellular Carcinoma Patients Using Unsupervised Machine Learning and Its Effect on Survival: A Population-Based Study, ACG 2024 Annual Scientific Meeting Abstracts. Philadelphia, PA: American College of Gastroenterology.