Deep representation learning for clustering longitudinal survival data from electronic health records

Area of Science:

• Computational biology
• Biomedical informatics
• Machine learning in healthcare

Background:

• Precision medicine necessitates identifying distinct patient subgroups for tailored treatments.
• Electronic health records (EHRs) offer vast potential for uncovering these subgroups using machine learning.
• Existing methods often struggle to capture complex interactions in diagnosis trajectories and risk events, leading to heterogeneous subgroups.

Purpose of the Study:

• To develop and evaluate VaDeSC-EHR, a transformer-based variational autoencoder, for clustering longitudinal survival data from EHRs.
• To address limitations in capturing complex patient data interactions for improved subgroup identification.
• To enhance the development of precision medicine strategies through more accurate patient stratification.

Main Methods:

• Implementation of VaDeSC-EHR, a novel transformer-based variational autoencoder architecture.
• Clustering of longitudinal survival data extracted from electronic health records.
• Validation using synthetic and real-world benchmark datasets with known cluster labels.

Main Results:

• VaDeSC-EHR demonstrated superior performance compared to baseline methods on benchmark datasets.
• The model successfully identified four distinct subgroups in Crohn's disease patients.
• These subgroups exhibited divergent diagnosis trajectories and risk profiles, highlighting clinically and genetically relevant factors.

Conclusions:

• VaDeSC-EHR effectively identifies patient subgroups with distinct clinical and molecular characteristics.
• The method improves upon existing approaches for analyzing complex EHR data.
• VaDeSC-EHR serves as a powerful tool for advancing precision medicine by enabling novel patient subgroup discovery.