Deep learning for clustering of multivariate clinical patient trajectories with missing values

Area of Science:

• Computational biology
• Machine learning
• Precision medicine

Background:

• Precision medicine necessitates patient stratification based on disease presentation for tailored treatments.
• Neurological disorders present complex stratification challenges due to multifactorial nature and time-series data with missing values.

Purpose of the Study:

• To address the challenge of clustering multivariate short time series with missing values for patient stratification.
• To introduce a novel deep learning-based method for improved patient subgroup identification.

Main Methods:

• Proposed variational deep embedding with recurrence (VaDER), a Gaussian mixture variational autoencoder framework.
• Extended the framework to model multivariate time series and handle missing data directly.
• Validated VaDER on simulated and benchmark datasets with varying degrees of missingness.

Main Results:

• Accurately recovered clusters from simulated and benchmark data, demonstrating robustness to missing values.
• Successfully stratified Alzheimer's disease and Parkinson's disease patients into distinct subgroups.
• Identified clinically divergent disease progression profiles within patient subgroups, reflecting known disease aspects.

Conclusions:

• VaDER offers a valuable tool for patient stratification and multivariate time-series clustering.
• The method shows promise for advancing precision medicine by enabling more accurate patient subgrouping.