0565 Unsupervised Cluster Analysis in Sleep Apnea Research: An Empirical Evaluation of Methodological Decisions

Abstract Introduction Cluster analysis is increasingly used in obstructive sleep apnea (OSA) research to identify clinically and biologically meaningful subtypes. A major gap in current OSA clustering research is the limited evaluation of how methodological choices affect the results. Here we used a large population-based cohort to perform an empirical evaluation of methodological decisions. Methods Utilizing data from the Hispanic Community Health Study/Study of Latinos, we applied clustering to 23 features from four domains: sleep timing/duration, sleep disturbance/sleepiness, heart rate during sleep, and sleep-disordered breathing, along with age, sex, and BMI. Analyses were performed in the full sample (n=14,512), moderate-to-severe OSA (respiratory event index, REI>=5; n=4,403) and severe OSA (REI>=15; n=1,623). We permutated combinations of supervised variable selection (LASSO and univariate regression predicting REI) and unsupervised dimension reduction methods (PCA and two-stage post-clustering selection), and of clustering algorithms (K-means, hierarchical clustering, latent profile analysis). We also performed “null model” comparisons, applying the entire clustering pipeline over datasets with all features permuted to nullify any correlations between them. Results from the null model were used to normalize the performance measures of observed data clustering. Performance measures included Davies-Bouldin Index (DBI), Calinski-Harabasz Index (CHI), Silhouette Score, and normalized Shannon’s entropy. These were aggregated to a composite rank score (CRS). Held-out samples were used for external evaluation via cluster-structure projection and decision-tree classifier. Additional assessments included robustness to imputation (adjusted RAND index) and visualization via t-SNE. Results K-means cluster showed the highest stability across imputed data (ARI: mean=0.81, SD=0.14) and consistently achieved the best performance when paired with PCA (CRS: mean=0.88, SD=0.03). Hierarchical clustering frequently failed to project cluster assignments in held-out data (13 out of 38 projections) and to distinguish real structure from noise (all metrics within null-model range in 19 out of 41 comparisons). Clustering implemented with PCA or univariate regression consistently identified two to three clusters and had moderate to high performance metrics. Conclusion Applying clustering algorithms over “null data” and assessment over a held-out dataset are useful for validating clustering results. While hierarchical clustering may be useful for exploratory analysis, K-means appears to offer a more robust default choice for routine analysis. Support (if any)