Feature Reduction or Sample Reduction? A Stability Analysis of Parkinson's Disease Clustering.

Area of Science:

• Neuroscience
• Computational Biology
• Biostatistics

Background:

• Clustering methods are vital for understanding Parkinson's disease (PD) phenotypic heterogeneity.
• However, the reproducibility of PD subtype solutions derived from clustering remains a significant challenge.

Purpose of the Study:

• To investigate the impact of feature and sample reduction on clustering stability in Parkinson's disease datasets.
• To determine whether reducing data features or patient samples has a greater effect on the reliability of PD subtype identification.

Main Methods:

• Utilized baseline data from the Parkinson's Progression Markers Initiative (PPMI) cohort.
• Applied K-means, Gaussian Mixture Models (GMM), and DBSCAN algorithms.
• Systematically reduced features and samples (40%, 60%, 80%, 100%) and assessed cluster stability using Adjusted Rand Index (ARI).

Main Results:

• Sample reduction demonstrated a more pronounced effect on agreement with reference solutions across all clustering methods compared to feature reduction.
• Feature reduction primarily impacted the run-to-run reproducibility of clustering solutions.
• K-means exhibited the highest robustness, while GMM and DBSCAN showed reduced reproducibility under feature reduction and altered noise assignment, respectively.

Conclusions:

• Cohort size, reflected by sample reduction, appears to be a more critical factor for reference-solution stability in PD clustering than the number of features.
• Findings highlight the sensitivity of clustering stability to data subsetting in PD research.
• Emphasizes the need for careful consideration of sample size and feature selection in studies of PD heterogeneity.