Consensus Clustering

The algorithm iteratively subsamples the original patient dataset and applies a base clustering algorithm (e.g., hierarchical or k-means) to each subsample. By analyzing how often pairs of patients co-cluster across these perturbed runs, it builds a consensus matrix that quantifies cluster membership stability. This process distinguishes robust biological signals from random noise, ensuring that identified subgroups are reproducible and not artifacts of a single algorithmic run.

The core output is a N×N consensus matrix, where each entry represents the proportion of resampling iterations in which patient i and patient j were assigned to the same cluster. Values range from 0 (never co-clustered) to 1 (always co-clustered). This matrix is then used as a similarity measure for final hierarchical clustering, and its visual inspection via a consensus heatmap reveals the natural number of stable subgroups without requiring a pre-specified k.

Unlike k-means, consensus clustering provides quantitative metrics to determine the optimal number of clusters (k). Analysts track the cumulative distribution function (CDF) of the consensus matrix and the proportion of ambiguous clustering (PAC) score across a range of k values. A lower PAC score indicates a flatter CDF and higher cluster stability, enabling data-driven selection of the true number of patient subtypes rather than relying on heuristic methods.

Consensus clustering generates item-consensus plots that display the mean consensus for each patient within their assigned cluster. Patients with low cluster consensus scores are identified as having uncertain or borderline membership, highlighting individuals who may represent transitional disease states or mixed phenotypes. This soft assignment capability provides clinicians with a measure of confidence in a patient's subtype classification, directly informing treatment ambiguity.

The consensus framework is a meta-algorithm that wraps around any base clustering method. Common implementations use hierarchical clustering, k-means, or partitioning around medoids (PAM) as the inner algorithm. This flexibility allows researchers to select the base method most appropriate for their data structure (e.g., Euclidean distance for continuous biomarkers, Gower distance for mixed clinical data) while still benefiting from the stability assessment of the consensus wrapper.

A powerful extension applies consensus clustering to multiple data types independently before fusing the results. For example, separate consensus matrices can be built from genomics, proteomics, and clinical data for the same patient cohort. These matrices are then integrated using Similarity Network Fusion (SNF) or averaged to produce a final multi-view consensus, ensuring that patient subgroups are robust across all molecular layers and not driven by a single assay's noise.

The foundational machine learning paradigm that groups patients based on inherent data similarities without predefined labels. Unlike supervised methods, these algorithms discover natural disease subtypes by optimizing internal criteria such as compactness or connectivity. Consensus clustering builds upon this by aggregating multiple unsupervised runs to overcome the instability of single solutions.

A centroid-based partitioning algorithm that minimizes within-cluster variance by iteratively updating cluster centers. Key characteristics:

• Requires pre-specification of k clusters
• Assumes spherical, equally-sized groups
• Sensitive to initialization and outliers Consensus clustering frequently uses K-Means as its base learner, resampling to determine the optimal k and mitigate sensitivity to random initialization.

Builds a dendrogram representing nested patient groupings through either agglomerative (bottom-up) or divisive (top-down) strategies. Unlike K-Means, it does not require a predefined cluster count. Consensus clustering often applies hierarchical methods to the consensus matrix itself, extracting robust partitions from the aggregated similarity structure.

A validation framework that quantifies the reproducibility of patient clusters under data perturbation. Methods include:

• Bootstrapping with resampled observations
• Subsampling features or patients
• Noise injection to test robustness Consensus clustering is itself a stability analysis technique, measuring how frequently pairs of patients co-cluster across resampled iterations to identify genuinely robust subgroups.

An internal cluster validation metric ranging from -1 to 1 that measures how similar a patient is to its own cluster versus neighboring clusters. Values near +1 indicate well-separated, compact clusters. In consensus clustering workflows, the silhouette score is often computed on the final consensus partition to quantitatively assess the quality of the identified patient subgroups.

A probabilistic clustering model that assumes data are generated from a mixture of Gaussian distributions, providing soft cluster assignments with membership probabilities. Unlike hard clustering methods, GMMs capture clinical ambiguity where a patient may belong to multiple subgroups. Consensus approaches can aggregate GMM outputs to produce robust probabilistic patient stratifications.