Unsupervised Clustering

A centroid-based algorithm that partitions patients into a predefined number (k) of clusters by minimizing the within-cluster sum of squares (WCSS). The algorithm iteratively assigns data points to the nearest centroid and recalculates centroids until convergence.

• Strengths: Computationally efficient and scales well to large datasets.
• Limitations: Assumes spherical clusters of equal size and requires pre-specifying k.
• Clinical Use Case: Grouping patients based on standardized lab values to identify broad metabolic profiles.

An algorithm that builds a tree-like structure (dendrogram) of patient relationships without requiring a pre-specified number of clusters. It operates via two approaches: agglomerative (bottom-up merging of individual points) or divisive (top-down splitting of the whole dataset).

• Key Output: A dendrogram that visualizes nested groupings at multiple resolutions.
• Distance Metrics: Relies on linkage criteria (Ward, complete, average) to define cluster proximity.
• Clinical Use Case: Analyzing tumor gene expression data to reveal a taxonomy of cancer subtypes.

A probabilistic model that assumes all data points are generated from a mixture of a finite number of Gaussian distributions. Unlike hard clustering, GMM provides soft assignments, quantifying the probability of a patient belonging to each cluster.

• Flexibility: Can model elliptical clusters of varying sizes and orientations.
• Covariance Types: Supports full, tied, diagonal, and spherical covariance matrices.
• Clinical Use Case: Identifying patient subgroups with overlapping symptoms where a patient might have partial membership in multiple disease endotypes.

A density-based algorithm that groups patients in high-density regions and identifies outliers as noise. It defines clusters as dense areas separated by sparse areas, requiring two parameters: epsilon (neighborhood radius) and minPts (minimum points to form a dense region).

• Key Advantage: Does not require a predefined cluster count and discovers arbitrarily shaped clusters.
• Noise Handling: Explicitly classifies outliers, which can represent rare disease presentations.
• Clinical Use Case: Detecting anomalous patient trajectories in longitudinal electronic health records.

An internal validation metric that measures how similar a patient is to its own cluster compared to other clusters. The score ranges from -1 to 1, where a high positive value indicates well-separated, compact clusters.

• Calculation: For each point, it compares the mean intra-cluster distance (a) to the mean nearest-cluster distance (b).
• Interpretation: Scores near 0 suggest overlapping clusters; negative scores indicate misclassification.
• Clinical Use Case: Objectively comparing clustering solutions with different k values to select the optimal number of patient subgroups.

A resampling-based methodology that aggregates results from multiple clustering runs on perturbed data subsets to identify robust and stable patient subgroups. It builds a consensus matrix quantifying how frequently pairs of patients are assigned to the same cluster.

• Process: Iteratively applies a base algorithm (e.g., K-Means or Hierarchical) to subsamples of the data.
• Output: A stability heatmap and a consensus dendrogram that reveal the most reproducible cluster structure.
• Clinical Use Case: Defining molecular subtypes in large-scale cancer genomics projects like The Cancer Genome Atlas (TCGA).