Survival analysis of patient groups defined by unsupervised machine learning clustering methods based on patient metabolomic data.

• DOI: 10.1016/j.csbj.2023.10.033
• Publication Date: 2023-10-19T04:16:55Z
• Authors: Caroline Bailleux; David Chardin; Jean-Marie Guigonis; Jean-Marc Ferrero; Yann Chateau; Olivier Humbert; Thierry Pourcher; Jocelyn Gal
• Keywords: Breast cancer; Survival; Proof-of-concept; [SDV.CAN]Life Sciences [q-bio]/Cancer; Untargeted; Unsupervised machine learning; Clustering; TP248.13-248.65; 3. Good health; [INFO.INFO-AI]Computer Science [cs]/Artificial Intelligence [cs.AI]; Biotechnology; Research Article

PurposeMeta-analyses failed to accurately identify patients with non-metastatic breast cancer who are likely benefit from chemotherapy, and metabolomics could provide new answers. In our previous published work, were clustered using five different unsupervised machine learning (ML) methods resulting in the identification of three clusters distinct clinical simulated survival data. The objective this study was evaluate outcomes, extended follow-up, same 5 learning.Experimental designForty-nine patients, diagnosed between 2013 2016, BC included retrospectively. Median follow-up 85.8 months. 449 metabolites extracted tumor resection samples by combined Liquid chromatography-mass spectrometry (LC–MS). Survival analyses reported grouping together Cluster 1 2 versus cluster 3. Bootstrap optimization applied.ResultsPCA k-means, K-sparse Spectral clustering most effective predict 2-year progression-free bootstrap (PFSb); as example, PCA k-means method, PFSb 94% for 1&2 82% 3 (p=0.01). method performed best, higher reproducibility (mean HR=2 (95%CI [1.4-2.7]); probability p≤0.05 85%). Cancer-specific (CSS) overall (OS) highlighted a discrepancy ML methods.ConclusionOur is proof-of-principle that it possible use on metabolomic data PFS best performance k-means. A larger population needed draw conclusions CSS OS analyses.