Patient Visit Trajectories

Using visit trajectory analysis to identify patient subgroups with distinct survival patterns

Overview

This project presents a novel approach for stratifying cancer patients using visit trajectory analysis. By leveraging electronic medical records (EMRs) from the Department of Veterans Affairs (VA), we developed a methodology to identify patient subgroups with distinct survival patterns based on their healthcare utilization over time.

Key Contributions

  • Trajectory-based patient representation: Developed a method to represent patient visits as 2D coordinates using ICD codes, enabling visualization and analysis of healthcare trajectories
  • Unsupervised stratification: Applied agglomerative hierarchical clustering to group patients based on their visit trajectory features
  • Survival analysis: Demonstrated that identified clusters show statistically significant differences in survival outcomes

Methodology

The approach consists of three main steps:

  1. Visit Mapping: Patient visits are mapped to 2D coordinates using a model trained on ICD-10-CM codes, where clusters of similar medical conditions form distinct regions in the embedding space

  2. Feature Extraction: For each patient, we compute features based on their trajectory through this embedded space, including:
    • Distance traveled between consecutive visits
    • Directional changes in the trajectory
    • Density of visits in different regions
  3. Clustering: Hierarchical clustering identifies patient subgroups with similar trajectory patterns

Results

Applied to a cohort of over 210,000 cancer patients from the VA healthcare system, the method identified distinct patient subgroups with significantly different survival curves (p < 0.001). The trajectory-based features captured clinically meaningful patterns related to disease progression and healthcare utilization.

Technologies

  • Python
  • Clustering
  • Electronic Medical Records processing

Publication

This work was presented at the CCS 2024