Unsupervised Clustering Successfully Predicts Prognosis in NSCLC Brain Metastasis Cohorts

Area of Science:

• Oncology
• Data Science
• Biostatistics

Background:

• Complex algorithms in computer-aided systems can burden clinicians.
• Simple, effective methods are needed for patient categorization.
• Unsupervised cluster analysis can identify prognostic subgroups in non-small-cell lung cancer (NSCLC) with brain metastasis (BM).

Purpose of the Study:

• To employ unsupervised cluster analysis for identifying prognostic subgroups of NSCLC patients with BM.
• To compare the effectiveness of two-step clustering (TSC) and hierarchical cluster analysis (HCA) in patient stratification.
• To assess if these clustering methods offer a viable alternative to traditional prognostic assessments.

Main Methods:

• Retrospective collection of data from 95 NSCLC patients with BM.
• Application of TSC and HCA to baseline clinical data for subgroup identification.
• Comparison of clustering results with Diagnosis-Specific Graded Prognostic Assessment (DS-GPA) and survival analysis using Kaplan-Meier and log-rank tests.

Main Results:

• Hierarchical cluster analysis (HCA) demonstrated the highest discriminatory power (C-index = 0.721), outperforming TSC (C-index = 0.650) and DS-GPA (C-index = 0.709).
• Both TSC and HCA significantly divided patients into prognostic clusters (p < 0.001).
• The clustering methods showed comparable prognostic performance to the DS-GPA index.

Conclusions:

• Unsupervised clustering methods (TSC and HCA) are comparable to the DS-GPA index for prognostic performance in NSCLC patients with BM.
• These data-driven clustering approaches offer a promising perspective for patient stratification.
• Further validation is required to establish the definitive role of unsupervised clustering in prognostic modeling for this patient group.