Background: Healthcare databases present large opportunities to study patient care pathways for specific disease. Each patient’s medical history is a timeline containing medical events (hospitalizations, consultations, drug delivery, etc.). Since each patient is unique, it is rare to find two patients with identical care sequences, and even less when studying thousands of patients. It has been shown that user-friendly and comprehensive data visualizations have a strong impact on the interpretation of study results. One of the challenges is to provide such appropriate visuals to ensure that results are well understood by the medical community, as well as patients and their relatives.
Objectives: This work aims at exploring and providing methodological recommendations on the use of innovative data visualizations supporting analytic methods, including artificial intelligences, to address healthcare pathway questions.
Methods: Several methods were tested and applied on lung cancer patients at different steps of their healthcare pathways, from diagnosis to treatment, up to death or end of follow-up. Several methodological steps were considered prior choosing the most appropriate analysis and visualization of healthcare pathway, depending on the study questions. The chosen approaches differ based on the type of variables (categorical or continues), as well as the number of simultaneously modelled variables.
Results: To visualize the different medical diagnostic tests and consultations performed prior to the initiation of a treatment for a cohort of patients with lung cancer, we considered a Sankey and sunburst diagram. It is particularly appropriate to represent pathways with a low diversity (only few care options at each step), and it explicitly visualizes the order of the different events (ex: consultation, followed by imaging investigation, followed by biopsies). However, these two options do not provide any temporality of the event sequences. As time to events is of major interest to study, for instance, the duration and sequences of the different treatment lines, a TAK® approach (Time sequence analysis through K-clustering) is extremely relevant. To highlight main patterns observed and main deviations, we can go further with process mining methods, particularly adapted to summarize complex, heterogenous and diverse pathways.
Conclusions: Our work resulted in a decision tree of methodologies that can be applied in order to analyze and visualize healthcare pathway depending on each clinical situation and study question in lung cancer. This could be applied to other types of cancer or even to other diseases.
