Clinical outcome-guided deep temporal clustering for disease progression subtyping

Area of Science:

• Biomedical Informatics
• Computational Biology
• Data Science

Background:

• Complex diseases show varied progression, requiring methods to identify subtypes for personalized medicine.
• Current approaches often lack clustering-specific representations or fail to integrate clinical outcomes, limiting utility.

Purpose of the Study:

• To develop a unified framework for subtyping longitudinal progressive diseases.
• To improve patient representation for clustering by integrating disease progression data and clinical outcomes.

Main Methods:

• Proposed Outcome-Guided Deep Temporal Clustering (OG-DTC) framework.
• Utilized a GRU-based seq2seq architecture for temporal dynamics, integrating k-means clustering and outcome regression.
• Clustered representations using Gaussian mixture models and validated through reproducibility, stability, and significance tests.

Main Results:

• Applied the framework to Alzheimer's Disease (AD) clinical trials, identifying three distinct subtypes.
• Subtypes showed unique progression patterns and differentiated clinical declines.
• Ablation studies confirmed the contribution of each model component, with joint optimization improving representations.

Conclusions:

• The temporal clustering framework provides robust subtyping for longitudinal diseases.
• The method has potential to capture subtype variability in clinical outcomes for personalized treatments.