VaDER

VaDER clusters multivariate short clinical time series (patient trajectories) using a Gaussian mixture variational autoencoder with recurrence to stratify patients and accommodate missing values for applications in precision medicine.

Key Features:

• Gaussian mixture variational autoencoder (VAE): Employs a Gaussian mixture VAE framework to model complex multivariate time series distributions.
• Recurrence modeling: Incorporates recurrence mechanisms to capture temporal dependencies in clinical trajectories.
• Handling missing values: Directly manages missing values within longitudinal clinical datasets.
• Clustering multivariate short time series: Designed specifically for clustering multivariate short time series characteristic of clinical patient trajectories.
• Validation on simulated and benchmark datasets: Demonstrated ability to recover ground truth clustering on simulated and benchmark datasets with varying degrees of missingness.

Scientific Applications:

• Alzheimer's disease patient stratification: Stratifies Alzheimer's disease patients into subgroups with distinct clinical progression profiles.
• Parkinson's disease patient stratification: Stratifies Parkinson's disease patients into subgroups with distinct clinical progression profiles.
• Precision medicine: Supports patient stratification to inform individualized treatment strategies based on divergent disease progression.
• Multivariate time-series clustering in medical research: Applicable to clustering longitudinal clinical data, including neurological disorders, in the presence of missing data.

Methodology:

Uses a Gaussian mixture variational autoencoder extended with recurrence mechanisms, directly handles missing values, and is validated on simulated and benchmark datasets with known ground truth clustering.