DBSCAN

DBSCAN defines clusters through two critical hyperparameters: epsilon (ε), the maximum radius of a neighborhood, and minPts, the minimum number of points required to form a dense region. A core point has at least minPts neighbors within its ε-radius. Border points fall within a core point's neighborhood but lack sufficient neighbors themselves. All other points are classified as noise or outliers. This density-reachability framework allows DBSCAN to discover clusters of arbitrary, non-spherical shapes—a critical advantage over centroid-based methods like K-Means when identifying irregular patient subgroups in multi-omics space.

A defining characteristic of DBSCAN is its explicit classification of points as noise rather than forcing every observation into a cluster. In clinical contexts, this is invaluable: patients who do not fit any molecular subtype are flagged as outliers rather than being incorrectly assigned. This property supports anomaly detection in biomarker studies, where outlier patients may represent rare disease variants, misdiagnosed cases, or novel endotypes worthy of further investigation. Unlike soft clustering methods that assign partial membership, DBSCAN provides a hard boundary that clearly delineates patients who genuinely belong to a subgroup from those who do not.

Unlike centroid-based algorithms that assume spherical clusters, DBSCAN can identify non-linear, elongated, and irregularly shaped patient groupings. This is critical in biological data where disease trajectories and molecular signatures often form complex manifolds rather than Gaussian distributions. For example, in single-cell RNA sequencing, cell populations frequently exhibit branching differentiation paths that DBSCAN can trace without distortion. The algorithm's ability to follow the natural topology of the data makes it particularly effective when combined with non-linear dimensionality reduction techniques like UMAP or t-SNE for visual validation of discovered subgroups.

The performance of DBSCAN is highly dependent on the selection of epsilon (ε) and minPts. A common heuristic sets minPts to twice the number of dimensions in the dataset, then uses a k-distance graph to identify the elbow point for ε. In patient stratification workflows, domain knowledge should guide these choices: too small an ε fragments true biological subgroups, while too large an ε merges distinct endotypes. HDBSCAN, a hierarchical extension, mitigates this sensitivity by automatically extracting clusters across varying density levels, making it a preferred alternative when cluster density varies significantly across the patient population.

DBSCAN achieves an average runtime complexity of O(n log n) when implemented with spatial indexing structures like R*-trees or kd-trees, making it suitable for moderately sized clinical datasets. However, it struggles with high-dimensional data due to the curse of dimensionality, where distance metrics lose discriminative power. In biomarker identification pipelines, this is addressed by applying dimensionality reduction (PCA, UMAP) before clustering. Additionally, DBSCAN assumes uniform density across clusters; datasets with widely varying cluster densities—common in heterogeneous patient populations—may require the HDBSCAN variant for robust stratification.

Validating DBSCAN-derived patient clusters requires both internal and external metrics. The Silhouette Score measures how well-separated clusters are, though it penalizes non-spherical shapes that DBSCAN naturally produces. Cluster stability analysis via resampling assesses reproducibility under data perturbation. For clinical relevance, discovered subgroups should be evaluated against ground-truth outcomes: survival curves (Kaplan-Meier analysis), differential treatment response, or enrichment of known biological pathways. A cluster with high statistical stability but no clinical interpretability offers limited translational value in precision medicine applications.

DBSCAN classifies every data point into three categories based on the minPts and epsilon parameters:

• Core points: Have at least minPts neighbors within radius epsilon; these form the dense heart of a cluster.
• Border points: Have fewer than minPts neighbors but fall within epsilon of a core point; they extend cluster boundaries.
• Noise points: Fail both conditions; these are outliers—patients who do not fit any defined subgroup, often clinically significant as atypical presentations or data quality issues.

A graph-based representation where nodes represent individual patients and edges encode clinical or molecular similarity. DBSCAN can be conceptualized as performing density-based community detection on an epsilon-neighborhood graph. When combined with Similarity Network Fusion (SNF), these networks integrate multi-omics data to identify robust patient subtypes that are consistent across genomic, proteomic, and clinical modalities.

An internal validation metric ranging from -1 to 1 that measures how similar a patient is to its own cluster versus other clusters. For DBSCAN, the silhouette score is particularly useful for tuning epsilon and minPts parameters. However, it penalizes noise points, so it should be interpreted carefully—a lower score may reflect valid outlier identification rather than poor clustering quality in clinical cohort analysis.

Endotypes are disease subtypes defined by distinct pathobiological mechanisms rather than observable symptoms. DBSCAN excels at endotype discovery because it:

• Does not force patients into clusters (unlike K-Means)
• Identifies arbitrarily shaped molecular signatures
• Flags patients with unique pathophysiology as noise for further investigation This makes it a preferred tool for identifying novel therapeutic targets in heterogeneous diseases like asthma, sepsis, and cancer.