Self-Organizing Maps (SOM)

The defining characteristic of a SOM is its ability to preserve the topological relationships of the input space. Similar input vectors map to nearby or identical nodes on the output grid, while dissimilar vectors map to distant nodes. This is achieved through a competitive learning process where the Best Matching Unit (BMU) and its topological neighbors are updated simultaneously, ensuring the map's spatial ordering reflects the data's intrinsic structure.

SOMs operate via a two-phase iterative process:

• Competition: For each input vector, the node with the smallest Euclidean distance (the BMU) is identified.
• Cooperation & Adaptation: The BMU and its neighbors within a defined radius are adjusted to move closer to the input vector. The learning rate and neighborhood radius decrease over time, transitioning the map from an initial ordering phase to a fine-tuning convergence phase.

A primary use case in patient stratification is the visualization of complex, high-dimensional omics data. The trained SOM grid can be rendered as a Unified Distance Matrix (U-Matrix), which color-codes the distances between adjacent nodes. Light areas represent cluster boundaries, while dark areas represent dense, similar patient groups, allowing researchers to visually identify natural disease subtypes without predefined labels.

Beyond cluster visualization, SOMs provide component planes—individual heatmaps of the grid for each input variable. By comparing component planes side-by-side, clinical data scientists can visually correlate features. For example, a plane showing high expression of a specific biomarker can be directly compared to a plane showing a high polygenic risk score, revealing multivariate relationships that drive patient subgroup separation.

Unlike hard clustering algorithms like K-Means, SOMs naturally model transitional states. Patients mapped to nodes on the border between two dense regions on a U-Matrix represent ambiguous or intermediate phenotypes. This is critical for modeling disease progression or continuous treatment responses, where a patient does not fit neatly into a discrete cluster but exists on a continuum between molecular subtypes.

SOMs inherently function as an anomaly detection system. Input vectors that activate nodes with a high quantization error (a large distance between the input and its BMU) are statistical outliers. In a clinical context, these patients represent rare disease variants, data entry errors, or novel molecular profiles that warrant further investigation, making SOMs a valuable tool for quality control and novel endotype discovery.

A machine learning technique for grouping patients based on inherent data similarities without predefined labels. SOMs perform a form of topology-preserving clustering where the spatial arrangement of nodes on the map reflects the similarity structure of the input data.

• Reveals natural disease subtypes from molecular data
• Unlike K-Means, SOM clusters maintain neighborhood relationships
• Essential for discovering novel patient endotypes

The process of reducing the number of random variables under consideration by obtaining a set of principal variables. SOMs achieve this by projecting high-dimensional patient data onto a discretized 2D grid while preserving the topological structure of the input space.

• Enables visual exploration of complex multi-omics datasets
• The U-Matrix visualization reveals cluster boundaries on the SOM grid
• Complements PCA, t-SNE, and UMAP for different analytical perspectives

A manifold learning technique that better preserves global data structure than t-SNE. While SOMs use a fixed grid topology, UMAP constructs a fuzzy topological representation of the data manifold.

• Often used alongside SOMs for validating patient subgroup structures
• Preserves both local and global neighborhood relationships
• Widely adopted for visualizing patient cohorts in single-cell studies

A resampling-based methodology that aggregates results from multiple clustering runs to identify robust and stable patient subgroups. SOMs can serve as the base clustering algorithm within consensus frameworks.

• Quantifies cluster stability through consensus matrices
• Addresses the sensitivity of SOM initialization to random weight assignment
• Essential for generating reproducible clinical stratifications

Graph-based representations where nodes are patients and edges represent clinical or molecular similarity. SOMs naturally produce a neighborhood graph where adjacent map units represent similar patient profiles.

• Enables community detection for stratification using algorithms like Louvain
• The SOM grid itself functions as a similarity-preserving network
• Bridges clustering and graph-based analytical approaches

A method for studying the shape of complex patient data using persistent homology. SOMs share the fundamental goal of preserving topological features—the connectedness and continuity of the data manifold—during dimensionality reduction.

• Detects high-dimensional voids and connectivity patterns
• The SOM's topology-preserving property is a direct application of topological principles
• Used to validate that SOM projections faithfully represent the underlying data structure