Clustering multiply imputed multivariate high-dimensional longitudinal profiles.

Area of Science:

• Statistics
• Data Science
• Biostatistics

Background:

• Functional data analysis often requires complete data for techniques like principal component analysis (PCA).
• Missing observations are common in longitudinal health studies, posing a challenge for standard analytical methods.

Purpose of the Study:

• To propose and validate a novel method for clustering multivariate functional data with missing observations.
• To address the challenge of dimension reduction techniques requiring complete data matrices in functional data analysis.

Main Methods:

• Multiple imputation was used to complete the dataset, followed by clustering each imputed dataset.
• Ensemble clustering was employed to synthesize partitions from imputed datasets, providing a robust final clustering.
• Uncertainty in cluster membership was quantified using ensemble agreement and consensus clustering fuzziness.

Main Results:

• The method was applied to heart failure (HF) patient data, utilizing daily measurements of four biomarkers.
• The algorithm identified a latent structure, dividing HF patients into two distinct clusters based on blood pressure and weight evolution.
• Data normalization (natural logarithm) and smoothing (cubic spline base) were applied to longitudinal outcomes.

Conclusions:

• The proposed method effectively clusters multivariate functional data with missing values, revealing underlying patient heterogeneity.
• The approach provides a reliable way to characterize cluster membership uncertainty in the presence of missing data.
• The identified clusters in HF patients suggest different disease progression patterns, aiding clinical stratification.