Clustering multivariate time series using Hidden Markov Models.

Area of Science:

• Data Science
• Biostatistics
• Health Informatics

Background:

• Multivariate time series with mixed categorical and continuous data are common in healthcare, representing individual health trajectories.
• Clustering these trajectories is challenging due to difficulties in defining meaningful distances between trajectories with categorical variables.

Purpose of the Study:

• To develop and evaluate an algorithm for clustering multivariate time series with mixed data types.
• To address the challenge of defining distances for trajectories containing categorical variables.

Main Methods:

• The proposed approach utilizes Hidden Markov Models (HMMs) to map each trajectory.
• A suitable distance metric is defined between HMMs.
• Clustering is performed using a distance matrix-based method.

Main Results:

• The algorithm was tested on simulated data, a synthetic validation set, and real-world data from the Health and Retirement Survey.
• The method demonstrated effectiveness in clustering trajectories with mixed data types.

Conclusions:

• The developed algorithm offers a practical solution for clustering multivariate time series with categorical variables.
• The method is implementable using standard R and Matlab packages, making it accessible to researchers without advanced statistical expertise.