Cluster Stability Analysis

A resampling-based methodology that aggregates results from multiple clustering runs to identify robust patient subgroups. The core mechanism involves:

• Repeatedly subsampling the original dataset (typically 80% of patients and features)
• Applying a base clustering algorithm (e.g., hierarchical clustering or k-means) to each subsample
• Constructing a consensus matrix where each entry represents the proportion of runs in which two patients co-clustered
• Applying hierarchical clustering to the consensus matrix to derive final stable assignments

The consensus matrix itself serves as a visual heatmap, where block-like diagonal structures indicate high stability. This method directly quantifies the robustness that cluster stability analysis seeks to measure.

An internal cluster validation metric that measures how similar a patient is to its own cluster compared to other clusters. The score for each point is calculated as:

• a(i): Mean distance to all other points in the same cluster (cohesion)
• b(i): Mean distance to points in the nearest neighboring cluster (separation)
• Silhouette s(i) = (b(i) - a(i)) / max{a(i), b(i)}

Values range from -1 to 1, where scores near 1 indicate well-separated, compact clusters, scores near 0 suggest overlapping clusters, and negative values indicate potential misclassification. The average silhouette width across all patients provides a global measure of clustering quality, making it a standard benchmark in stability analysis workflows.

A probabilistic model that assumes all data points are generated from a mixture of a finite number of Gaussian distributions. Unlike hard clustering methods, GMMs provide soft cluster assignments—each patient receives a posterior probability of belonging to each cluster.

Key properties relevant to stability analysis:

• The Expectation-Maximization (EM) algorithm iteratively estimates parameters
• Model selection uses criteria like Bayesian Information Criterion (BIC) to determine the optimal number of components
• Soft assignments allow uncertainty quantification, revealing patients who sit at cluster boundaries
• Stability can be assessed by comparing posterior distributions across bootstrap samples

This probabilistic framework naturally complements stability analysis by explicitly modeling the ambiguity in patient assignments.

A set-based similarity metric frequently used to quantify the agreement between two clustering solutions. For two clusters A and B, the Jaccard coefficient is defined as:

• J(A, B) = |A ∩ B| / |A ∪ B|

In cluster stability analysis, this metric serves several critical functions:

• Comparing clusters derived from bootstrap samples to original clusters
• Measuring the recovery rate of known patient subgroups under data perturbation
• Constructing cluster stability heatmaps that visualize pairwise cluster overlap across resampling iterations
• Thresholding Jaccard values (e.g., > 0.75) to declare clusters as stable and reproducible

Unlike correlation-based measures, Jaccard directly evaluates set membership agreement, making it particularly intuitive for clinical validation of patient subgroups.

A chance-corrected external validation metric that measures the similarity between two data clusterings. The ARI adjusts the Rand Index to account for the expected similarity by random chance:

• ARI = (RI - Expected_RI) / (max(RI) - Expected_RI)
• Values range from -1 to 1, where 1 indicates perfect agreement, 0 represents random labeling, and negative values suggest less agreement than expected by chance

In stability analysis, ARI is computed between:

• Clusters derived from full dataset vs. clusters from perturbed or subsampled data
• Multiple independent clustering runs with different initializations
• Clusters from different data modalities (e.g., genomics vs. proteomics)

ARI provides a rigorous, interpretable benchmark for declaring patient clusters as stable and reproducible across analytical variations.

A resampling technique that forms the statistical backbone of cluster stability analysis. The procedure involves:

• Drawing B bootstrap samples (typically 100-1000) from the original patient dataset with replacement
• Applying the identical clustering pipeline to each bootstrap sample independently
• Computing a stability metric (e.g., Jaccard, ARI, or consensus matrix values) between each bootstrap solution and the original clustering
• Aggregating results into stability distributions with confidence intervals

Key considerations include:

• Sample size sensitivity: Small clusters may dissolve under resampling, revealing instability
• Feature perturbation: Bootstrapping can be applied to features as well as patients
• Visualization: Stability paths plot cluster recovery rates across varying numbers of clusters (k)

This framework transforms clustering from a point estimate into a distribution, enabling rigorous statistical inference about subgroup reproducibility.