TreeScan™: A Novel Data-Mining Tool for Medical Product Safety Surveillance

Basic Details
Friday, August 25, 2017

Background: The tree-based scan statistic -- operationalized in freely available TreeScan™ software -- anchors a statistical signal detection approach to evaluate unexpected potential associations between exposures and outcomes of interest in electronic healthcare data. The approach takes advantage of the hierarchical nature of clinical concepts, including clinical outcomes and medical product exposures. Investigators can collect and analyze data on several thousand outcomes, or exposures simultaneously, while formally controlling for multiple hypothesis testing.

Objectives: This workshop will teach the theory of tree-based scan statistics, describe recent efforts by the U.S. FDA in using TreeScan™ for routine safety surveillance, demonstrate how to use TreeScan™ software, and then allow audience members to detect signals that have created by manipulating a simulated dataset. Investigators that seek to apply the TreeScan™ method to their own database sources will benefit from attending.

Description: The workshop agenda follows:

  1. Overview of tree-based scan statistics theory
  2. Rationale and findings of completed activities 
    • Vaccine Safety Surveillance: We will review prior work on exposure-based surveillance including scanning a tree of outcomes for associations with measles-mumps-rubella (MMR) vaccine, measles-mumps-rubella-varicella (MMRV) vaccine, and quadrivalent human papillomavirus (HPV4) vaccine. 
    • Drug Safety Surveillance: We will review prior work on outcome-based surveillance including scanning a tree of exposures for associations with angioedema.
  3. Discussion and questions from the audience
  4. Interactive demonstration of TreeScan™ software with audience participation. Audience members will have downloaded TreeScan™ to their laptops along with simulated datasets, and then will execute TreeScan™ on those datasets.
  5. Interactive signal detection exercise
    • Audience members will manipulate an outcome in the dataset by adding extra cases, and then will report on their ability to detect their injected signal. 
  6. Discussion and questions from the audience

Judith Maro, Rima Izem, Martin Kulldorff, Azadeh Shoaibi