The analysis of patient care trajectories, which represent the sequence of healthcare states experienced by individuals over time, has gained significant importance in modern healthcare research. These trajectories provide valuable insights into disease progression and treatment effectiveness. However, analysing patient care trajectories presents several challenges. Their sequential nature, where the order of states matters, necessitates specialized analytical methods. Furthermore, the heterogeneity of patient experiences and the complexity of healthcare systems contribute to the difficulty of identifying meaningful patterns and subgroups within these trajectories. Different methods providing valuable frameworks for sequence analysis have been proposed and implemented (e.g. TraMineR, Eventflow). However, they may not fully address the complexities of patient care trajectories. Our objective is to develop an original clustering methodology combining univariate and multivariate techniques to effectively segment patients based on their care trajectories. Such hybrid approach will leverage the strengths of both methods to identify meaningful subgroups within the patient population.

We used data from a publicly available longitudinal cohort of 2929 patients diagnosed with chronic HIV infection. Patient care pathways were tracked monthly from the time of diagnosis. Notably, the observation period varies across individuals due to differences in diagnosis dates. Both the univariate and the multivariate analysis employed Hierarchical Ascendant Clustering (HAC) as the primary clustering technique. For the univariate approach, HAC incorporated Optimal Matching algorithm to measure the string-based distance considering the order of states within each patient's care trajectories, thus enhancing the sensitivity of the clustering to temporal patterns. Alternatively, the distance matrix is computed using Dynamic Time Warping (DTW), a flexible metric that allows for non-linear alignment of time series, making it suitable for comparing trajectories with varying lengths and temporal patterns. For the multivariate approach, the core distance metric, DTW, was employed in conjunction with one specific variant, the Triangle Inequality and Point Clustering DTW (TC-DTW), which leverage the strengths of DTW while improving computational efficiency. The proposed methodology has been implemented in a user-friendly Python package named Trajectory Clustering Analysis (TCA).

Univariate clustering identified distinct patient groups based on individual care components, such as length of stay in each state measured in months. Multivariate analysis using DTW revealed more nuanced patient trajectories, particularly for patients undergoing concurrent treatments. These clusters were associated with varying patient characteristics. The TCA package provides functions for healthcare pathway analysis, including user-friendly visualizations and effective performance measures.

We proposed an approach for clustering states sequences, implemented within a user-friendly Python package. Inspired by time series clustering techniques, our method leverages the advantages of timed sequences, such as handling co-occurring events, irregular time sampling, and sparse events. This research has significant potential for analyzing patient care pathways, enabling a deeper understanding of patient experiences and supporting the development of more personalized healthcare interventions.