Simulation-derived best practices for clustering clinical data.

Area of Science:

• Clinical informatics
• Data science in healthcare
• Biostatistics

Background:

• Clustering analyses are vital for understanding patient phenotypes and disease trajectories in clinical medicine.
• Ensuring rigor, validity, and reproducibility in clinical clustering solutions is an ongoing challenge.
• Best practices for dissimilarity matrix calculation and clustering on mixed-type clinical data require evaluation.

Purpose of the Study:

• To evaluate best practices for dissimilarity matrix calculation and clustering on mixed-type clinical data.
• To compare the performance of various distance metrics and clustering algorithms on simulated and real-world clinical datasets.
• To identify optimal methods for enhancing patient subclassification and precision medicine.

Main Methods:

• Simulated clinical data (binary, continuous, categorical, and mixtures) were used to test 5 single distance metrics and 3 mixed distance metrics.
• Clustering was performed using hierarchical clustering (HC), k-medoids, and self-organizing maps (SOM).
• Performance was validated using Adjusted Rand Index (ARI) and silhouette width (SW) on simulated and two real-world datasets (chronic lymphocytic leukemia and intensive care unit admissions).

Main Results:

• Hierarchical clustering (HC) demonstrated superior performance over k-medoids and SOM, evidenced by higher ARI across data types.
• The DAISY mixed-type distance metric yielded the highest mean ARI for most mixed data types.
• DAISY combined with HC identified superior, separable clusters in both real-world clinical datasets.

Conclusions:

• The selection of appropriate mixed-type distance metrics is essential for optimal patient cluster separation and data utilization.
• Advanced metrics capable of handling multiple data types enhance the subclassification of diseases.
• Improved disease subclassification facilitates targeted treatments, precision medicine, clinical decision support, and better patient outcomes.