Head-to-head comparison of clustering methods for heterogeneous data: a simulation-driven benchmark.

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• Clustering mixed data (continuous and categorical variables) presents challenges.
• Selecting appropriate unsupervised machine learning methods is crucial for accurate analysis.

Purpose of the Study:

• To benchmark clustering strategies for mixed data.
• To compare model-based and distance-based methods using simulated and real-world clinical trial data.

Main Methods:

• Evaluated 4 model-based (Kamila, Latent Class Analysis, Latent Class Model [LCM], Clustering by Mixture Modeling) and 5 distance/dissimilarity-based methods (K-prototypes, Gower distance, Unsupervised Extra Trees dissimilarity with hierarchical clustering or Partitioning Around Medoids).
• Assessed performance using Adjusted Rand Index (ARI) on simulated data across 7 scenarios.
• Applied methods to EPHESUS heart failure clinical trial data.

Main Results:

• K-prototypes, Kamila, and LCM models demonstrated superior performance.
• Model-based methods generally outperformed dissimilarity-based methods (Partitioning Around Medoids, Hierarchical Clustering).
• LCM showed promise in clinical data for profile differentiation, prognosis (C-index), and identifying treatment benefit subgroups.

Conclusions:

• Significant performance differences exist among clustering algorithms for mixed data.
• Model-based methods (Kamila, LCM) and K-prototypes are recommended for heterogeneous data analysis.
• LCM offers valuable insights in clinical settings for patient subgroup identification and prognosis.