592P - Survival analysis of surgical management options for goblet cell adenocarcinoma: Insights from machine learning clustering

Background

Debate surrounds the surgical management of Goblet cell adenocarcinoma (GCA). However, right hemicolectomy (RHC) remains the standard surgical choice for GCA.

Methods

1154 and 867 patients with GCA were extracted from U.S. Surveillance, Epidemiology, and End Results (SEER) and National Cancer Registration and Analysis Service (NCRAS) databases respectively after excluding patients with missing data. Two clusters for each cohort were created using an unsupervised machine learning K-means model (cluster 0, cluster 1). Clustering was evaluated using silhouette (S), Davies Boulden (DB) and Caliniski Harabasz (CH) scores. Clusters were visualised using principal component analysis (PCA) and T-distributed stochastic neighbour embedding (T-SNE). SHapley Additive exPlanations (SHAP) identified clinical features that were the most useful in clustering. Kaplan Meier (KM) statistics and graphs were generated. Decision trees were developed to cluster individual patients in each of the two clusters.

Results

PCA and T-SNE visualisations showed distinct clusters with some overlap in both cohorts. SHAP revealed tumour size and stage as important variables for clustering in SEER, while age and stage were key variables in NCRAS. Clusters description and survival are shown in the table. The two clusters were statistically different in survival based on Cox regression, Log-rank test and KM statistics in both cohorts. Appendectomy and RHC showed similar survival rates across both cohorts and within each cluster of both cohorts. Table: 592P

Characteristics of clusters and 5-year survival rates

Conclusions

After categorising cases into two clusters using unsupervised machine learning, RHC and appendectomy demonstrated comparable survival rates among patients with GCA. Machine learning could facilitate clinicians in exploring the impact of various treatment approaches on survival outcomes.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Hampshire Hospitals NHS Foundation Trust.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.