Targeted Learning for Variable Importance in Precision Medicine

• Activity Number: 676 - Analysis and Reporting: Benefit-Risk and Robust Models
• Type: Contributed
• Date/Time: Thursday, August 2, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Medical Devices and Diagnostics
• Author(s): Yue You* and Alan Hubbard and Rachael Callcut and Lucy Kornblith and Sabrinah Christie
• Companies: Division of Biostatistics, University of California, Berkeley and Division of Biostatistics, University of California, Berkeley and Zuckerberg San Francisco General Hospital, University of California and Zuckerberg San Francisco General Hospital, UCSF and Zuckerberg San Francisco General Hospital, UCSF
• Keywords: variable importance; targeted learning; trauma; causal inference; machine learning; precise medicine

Abstract

We proposed a parameter within a non-parametric model that measures the variable importance (VI) as the amount of attribution of that variable towards changes in the mean outcome. Specifically, for each of the candidate competing causes of the outcome, we utilized an estimate of this attribution using an approach based upon a combination of machine learning and causal inference via Targeted Learning. This approach allows for VI comparisons at the same scale, estimation not dependent on arbitrary parametric assumptions, and asymptotically linear (locally efficient) estimator where robust asymptotic inference is available. We implemented this approach to determine the VI of clinical measures in trauma patients in predicting the probability of mortality at different time periods from time of injury using data from three independent trauma cohorts. This approach allowed comparisons of VI within and between trauma cohorts and identified variables with the biggest potential "intervention" impact for mortality. Our results showed that the most important variables across all time intervals were mainly coagulation measures (such as initial International Normalized Ratio).