Predicting Health Outcomes from High-Dimensional Longitudinal Health Histories Using Relational Random Forests

• Activity Number: 637
• Type: Topic Contributed
• Date/Time: Thursday, August 13, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Zach Shahn* and David Madigan and Patrick Ryan
• Companies: Columbia University and Columbia University and Observational Health Data Sciences and Informatics
• Keywords: classification ; random forest ; labeled graph ; longitudinal ; high dimensional ; health histories

Abstract

With the rise of Electronic Health Records (EHRs), there is increasing interest in algorithms that predict health outcomes from health histories. However, it is challenging to translate all the potentially relevant information contained in a health history into a format that would be manageable for traditional machine learning algorithms. Specifically, it would be desirable to incorporate relative temporal relationships among multiple health events, a task made difficult by the vastness of the space of such predictors. To this end, we modify and repurpose a predictive modeling method (referred to in the remainder as Random Relational Forest, or RRF) originally developed in the context of speech recognition. RRF generates informative labeled graphs representing temporal relations among health events at the nodes of randomized decision trees. We illustrate the utility of RRF by predicting strokes in patients with prior diagnoses of Atrial Fibrillation (afib) and demonstrate potentially clinically meaningful improvements over more traditional predictive models.