About ParcoursVis

The Aviz Team and the French Social Security (CNAMTS) are interested in evaluating patient pathways for treatments. We are currently working on two main projects: (1) ParcoursVis and (2) ParcoursUrge. In ParcoursVis, we want to understand the differences between how treatments are de facto given at various sites and how they align with our conventional recommendations. In ParcoursUrge, we visualize the events that take place in the Emergency services of the Parisian hospitals to help doctors optimize patient management.

Both ParcoursVis and ParcoursUrge projects present us with several challenges, such as finding methodologies to effectively process large-scale data (multiple millions of patients with hundreds or thousands of events per patient), transforming raw prescription dataset into an intuitive and effective visualization, and recommending information tailored to the doctors’ needs ad-hoc to augment the explanation of the data for doctors.

These works are being done with the Health-Data-Hub (https://www.health-data-hub.fr/), and AP-HP (https://www.aphp.fr/). The project originally started as a collaboration with CMAP/École Polytechnique and CNAMTS.

The Data

For the moment, we focused on ParcoursVis on the adenoma of the prostate. With this focus, we extracted with our domain expert users the following high-level type of events (e.g., treatments) the patient follows in their care pathways. We do not yet consider which drugs the patients took, but we considered which molecules they were exposed to:

  • Alphabloc
  • Phyto
  • Surgery
  • Fivealpha
  • Interruption -- Patients interrupted their treatment but still benefit from them. We consider an interruption event when patient did not have any other event (related to our use-case) in the following 180 days.
  • No treatment -- We consider that patients are under no treatment when they did not have any other event (related to our use-case) in the following 360 days (~1 year).
  • Death

We also consider "composed types" of events when patients would benefit from multiple molecules at the same time, resulting in new treatments that the medical field recognized:

  • Alphabloc + Phyto
  • Fivealpha + Alphabloc
  • Fivealpha + Alphabloc + Phyto

The demo of ParcoursVis relies on synthetical data that we generated from a real cohort. This synthethical data tries to follow the same statistical distribution the real dataset has. However, all the "patients" our synthetical data has do not exist, which ensure anonymization from our real pool of patients. This synthethical dataset contains:

  • 2,000,000 patients
  • 50,640,284 low-level events (what the patients was reimbursed from)
  • 2,719,086 high-level events (what treatment did the patients follow. A high-level event, e.g., alphabloc, may come from patients buying multiple alphabloc pills (low-level events))