Patient Stratification

Stratification moves beyond traditional histology to define disease by its molecular mechanism rather than its anatomical site. This process, known as endotype discovery, identifies subgroups defined by a specific functional or pathobiological pathway.

• Distinguishes between patients who appear clinically similar but have distinct molecular drivers.
• Enables the matching of targeted therapies to the specific biological mechanism active in a patient subgroup.
• Relies on unsupervised learning to reveal natural structures in high-dimensional omics data without imposing prior clinical biases.

Patient data often contains tens of thousands of features (genes, proteins, metabolites) across relatively few samples. Stratification requires dimensionality reduction to extract the latent structure governing patient similarity.

• Principal Component Analysis (PCA) provides a linear projection that preserves global variance.
• t-SNE and UMAP offer non-linear manifold learning for visual exploration, with UMAP better preserving global data topology.
• The choice of reduction technique directly impacts which patient subgroups are revealed and must be validated against clinical outcomes.

Robust stratification demands the fusion of heterogeneous data types—genomic, transcriptomic, proteomic, and clinical—into a unified patient representation. Similarity Network Fusion (SNF) constructs a comprehensive patient similarity network by iteratively fusing individual data-type networks.

• Multi-Omics Factor Analysis (MOFA) discovers the principal sources of biological variation across multiple assays simultaneously.
• Integration avoids the pitfall of identifying clusters that are artifacts of a single data modality rather than true biological subgroups.

A discovered subgroup is only clinically useful if it is reproducible. Cluster stability analysis assesses robustness by measuring how consistently patients co-cluster under data perturbation or resampling.

• The Silhouette Score quantifies internal cohesion and separation, ranging from -1 to 1.
• Consensus clustering aggregates results from multiple algorithm runs to identify subgroups that persist across methodological variations.
• Validation against an independent external cohort is the gold standard for confirming that a stratification is not overfitting noise.

Not all patient heterogeneity is categorical. Trajectory inference orders patients along a continuous pseudotemporal path based on molecular profiles, modeling disease as a dynamic spectrum rather than static bins.

• Captures transitional states and disease progression gradients that discrete clustering obscures.
• Hidden Markov Models (HMM) model probabilistic transitions between unobserved health states in longitudinal data.
• This approach is critical for neurodegenerative and oncological conditions where disease evolves along a continuum.

Real-world patients often exhibit characteristics of multiple subgroups. Fuzzy C-Means and Gaussian Mixture Models (GMM) provide soft cluster assignments, giving each patient a probability vector of membership across all subgroups.

• Reflects clinical reality where disease boundaries are blurred rather than sharp.
• Enables risk calculations that incorporate uncertainty rather than forcing a hard binary classification.
• Supports clinical decision-making by quantifying the confidence of subgroup assignment for an individual patient.

A machine learning technique for grouping patients based on inherent data similarities without predefined labels, revealing natural disease subtypes. Key algorithms include:

• K-Means: Partitions patients into a predefined number of clusters by minimizing within-cluster variance.
• Hierarchical Clustering: Builds a dendrogram without requiring a pre-specified group count.
• DBSCAN: A density-based approach that identifies outliers as noise, robust to arbitrary cluster shapes.

The process of reducing the number of random variables under consideration by obtaining a set of principal variables, essential for visualizing high-dimensional patient data. Core techniques:

• Principal Component Analysis (PCA): A linear transformation to orthogonal components ordered by explained variance.
• t-SNE: A non-linear method optimized for preserving local structure in 2D/3D visualizations.
• UMAP: A manifold learning technique that better preserves global data structure and scales efficiently to large cohorts.

Graph-based representations where nodes are patients and edges represent clinical or molecular similarity. These networks enable community detection for stratification:

• The Louvain Algorithm maximizes modularity to identify densely connected patient groups.
• Similarity Network Fusion (SNF) integrates diverse data types by constructing and fusing individual similarity networks into a single comprehensive view, robustly capturing complementary information across omics layers.

A critical framework for assessing the robustness and clinical relevance of identified subgroups. Key metrics include:

• Silhouette Score: Measures how similar a patient is to its own cluster versus others, ranging from -1 to 1.
• Cluster Stability Analysis: Assesses reproducibility under data perturbation or resampling.
• Consensus Clustering: Aggregates results from multiple runs to identify robust, stable patient subgroups that are not artifacts of a single algorithm.

The process of identifying distinct disease subtypes defined by a specific functional or pathobiological mechanism, moving beyond clinical phenotype to molecular cause. Unlike phenotype-driven stratification, endotypes are linked to distinct therapeutic responses. This requires integrating multi-omics data with deep phenotyping to uncover the causal molecular pathways driving each subgroup, enabling true precision medicine.

A computational method that orders patients or cells along a continuous path based on molecular profiles, modeling dynamic disease progression rather than discrete clusters. Unlike static clustering, trajectory inference captures transitional states and branching differentiation. Commonly used in single-cell sequencing to map developmental lineages and in longitudinal cohort studies to model chronic disease evolution over time.