A self-supervised learning-based approach to clustering multivariate time-series data with missing values...

Area of Science:

• Machine Learning
• Biomedical Informatics
• Data Science

Background:

• Clustering multivariate time-series data is crucial for uncovering patterns in complex datasets.
• Existing methods struggle with missing values, requiring imputation that can introduce errors and computational overhead.
• Self-supervised learning offers a promising avenue for robust time-series analysis.

Purpose of the Study:

• To develop a self-supervised learning approach for clustering multivariate time-series data with inherent missing values.
• To address the limitations of imputation-based methods in time-series clustering.
• To identify distinct patient phenotypes in Traumatic Brain Injury (TBI) using time-series clinical data.

Main Methods:

• Introduced SLAC-Time, a Transformer-based clustering method utilizing time-series forecasting as a proxy task.
• Employed a joint learning approach for neural network parameters and cluster assignments.
• Iteratively clustered representations using K-means and updated parameters with pseudo-labels.

Main Results:

• SLAC-Time demonstrated superior performance over baseline K-means clustering based on multiple evaluation metrics.
• Successfully identified three distinct Traumatic Brain Injury (TBI) patient phenotypes.
• Phenotypes showed significant differences in clinical variables, outcomes (GOSE, ICU stay), and mortality.

Conclusions:

• SLAC-Time effectively clusters multivariate time-series data with missing values, outperforming traditional methods.
• The identified TBI phenotypes offer valuable insights for personalized medicine.
• These findings can guide the development of targeted clinical trials and therapeutic interventions for TBI patients.