A comparative analysis of unsupervised machine-learning methods in PSG-related phenotyping.

Area of Science:

• Sleep Medicine
• Computational Biology
• Data Science

Background:

• Obstructive sleep apnea (OSA) is a complex sleep disorder with diverse patient phenotypes.
• Previous studies used cluster analysis to identify OSA subtypes, but method selection's impact is unclear.
• Understanding OSA heterogeneity is crucial for personalized treatment strategies.

Purpose of the Study:

• To investigate how different clustering methods affect the formation and characteristics of OSA phenotypic clusters.
• To compare the performance and agreement of four clustering methods: Agglomerative Hierarchical Clustering, K-means, Fuzzy C-means, and Gaussian Mixture Model.
• To evaluate the suitability of soft clustering methods for OSA phenotyping.

Main Methods:

• Applied four clustering algorithms (Agglomerative Hierarchical Clustering, K-means, Fuzzy C-means, GMM) to 865 suspected OSA patients.
• Generated five clusters per method and analyzed physiological differences.
• Used visualization, Cohen's kappa statistics, and performance metrics to assess cluster similarity and quality.

Main Results:

• Two distinct OSA clusters were consistently identified across all four methods.
• Three other clusters showed overlapping features, varying by method.
• Fuzzy C-means and K-means exhibited the highest agreement (κ=0.87); AHC and GMM showed the lowest (κ=0.51).
• K-means performed best overall, with Fuzzy C-means showing strong performance and superior handling of overlapping clusters.

Conclusions:

• Clustering method selection critically influences OSA cluster identification and their physiological profiles.
• Soft clustering methods, especially Fuzzy C-means, offer significant advantages for nuanced OSA phenotyping.
• Findings underscore the need to consider algorithmic choices in OSA subtype research.