On an ensemble algorithm for clustering cancer patient data.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• The TNM staging system, based on Tumor, Node, Metastasis, is insufficient as cancer is not purely anatomic.
• Expanding TNM with new prognostic factors is crucial for accurate cancer patient outcome estimation.
• The ensemble algorithm for clustering cancer data (EACCD) aims to enhance TNM without altering its core definitions.

Purpose of the Study:

• To analyze the ensemble algorithm for clustering cancer data (EACCD) using a large breast cancer patient dataset.
• To compare EACCD's output with survival curves and evaluate the impact of its various settings.
• To assess EACCD's performance against alternative clustering approaches.

Main Methods:

• Utilized a large breast cancer patient dataset for EACCD analysis.
• Generated dendrograms to visualize relationships between patient survival curves.
• Investigated the influence of different EACCD parameters, including linkage functions and statistical tests.
• Evaluated the necessity of the dissimilarity learning step within EACCD.
• Compared EACCD with other clustering techniques and partitioning methods.

Main Results:

• EACCD generated robust dendrograms reflecting patient survival curve relationships, particularly with large numbers of runs (m).
• Dendrogram robustness is influenced by linkage functions but minimally by statistical tests for large m.
• Omitting the dissimilarity learning step in EACCD can degrade performance.
• Clustering solely on prognostic factors or using direct partitioning methods may yield misleading results.

Conclusions:

• EACCD effectively clusters cancer patient data when utilizing the Partitioning Around Medoids (PAM) algorithm for dissimilarity learning with sufficient runs (large m).
• Robust dendrograms are achievable with large m values in EACCD.
• The study validates EACCD as a valuable tool for cancer data analysis and outcome prediction.