Events Spring 2015

Targeted Learning: Applications in Precision Medicine

Feb 27, 2015 2:00 pm – 3:30 pm 

Add to Calendar


Mark van der Laan (UC Berkeley)


Calvin Lab auditorium

This talk will review Targeted Learning, a general statistical approach that uses machine learning to answer a user-supplied (low-dimensional) question of interest about the data generating distribution, while still providing statistical inference in terms of confidence intervals and p-values. Specifically, I will present targeted maximum likelihood estimation (TMLE), a two stage approach that 1) constructs an initial estimator using an optimal ensemble machine learning method called super learning, and 2) subsequently applies a targeted maximum likelihood updating step, involving maximum likelihood estimation along a least favorable parametric submodel through the initial estimator. The resulting TMLE of the target estimand is asymptotically normally distributed and efficient under regularity conditions, allowing for formal statistical inference.

Targeted Learning provides a powerful tool for developing and evaluating optimal dynamic treatment regimes, a central objective in precision medicine. A dynamic treatment regime is a rule for assigning and modifying a patient’s treatment over time in response to that patient’s individual evolving characteristics. In other words, a dynamic regime is a decision rule for delivering personalized care. Targeted Learning can be used to both learn the optimal dynamic regime (the decision rule that would optimize some outcome if applied to the target population as a whole) and to provide rigorous statistical inference for the comparative effectiveness of this learned optimal regime (as compared, for example, to simpler or non-personalized alternatives). Specifically, I will demonstrate a super learner of the optimal dynamic treatment regime itself and a TMLE of the mean counterfactual outcome under the optimal dynamic regime. I will also discuss extensions of this approach for learning and evaluating optimal dynamic treatments under resource constraints.

Two ongoing applied collaborations provide illustration. The first aims to learn the optimal rule for assigning blood-transfusions to trauma patients using observational data from multiple trauma centers. The second aims to learn the optimal rule for assigning behavioral interventions and incentives in order to retain patients in HIV care based on data from a sequentially randomized trial in Kenya.

Light refreshments will be served before the lecture at 1:30 p.m.