A Localized Prediction Algorithm for Personalized Trauma Care

• Activity Number: 308
• Type: Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Health Policy Statistics Section
• Author(s): Sara Moore* and Alan Hubbard and Mitchell J. Cohen
• Companies: and UC Berkeley and UC San Francisco
• Keywords: Personalized medicine ; Prediction ; Semiparametric methods ; Ensemble learning

Abstract

Traumatic injury is responsible for ten percent of deaths worldwide and remains the leading cause of death in Americans up to 45 years of age. Physicians providing emergency and trauma care often have limited time and information to make life-saving treatment decisions for severely injured patients. These decisions are typically made univariately or based on rudimentary multivariate scoring systems with limited predictive power for clinical outcomes of interest. We believe that trauma care would greatly benefit from improved ``personalized medicine'' decision algorithms. To this end, we propose a supervised classification method which performs both dimension and instance reduction data-adaptively to hone in only the most relevant information for a given patient's predictions. In both the feature selection and localized prediction steps, the algorithm utilizes an ensemble learner, the SuperLearner, which chooses the best combination of individual learners via minimization of cross-validated loss. Performance is demonstrated on both simulated and real-world patient data. Future applications include real-time treatment decision support for use by critical care clinicians.