Deep advantage learning for optimal dynamic treatment regime

Recently deep learning has successfully achieved state-of-the-art performance on many difficult tasks. Deep neural networks allow for model flexibility and process features without the need of domain knowledge. Advantage learning (A-learning) is a popular method in dynamic treatment regime (DTR). It models the advantage function, which is of direct relevance to optimal treatment decision. No assumptions on baseline function are made. However, there is a paucity of literature on deep A-learning. In this paper, we present a deep A-learning approach to estimate optimal DTR. We use an inverse probability weighting method to estimate the difference between potential outcomes. Parameter sharing of convolutional neural networks (CNN) greatly reduces the amount of parameters in neural networks, which allows for high scalability. Convexified convolutional neural networks (CCNN) relax the constraints of CNN for optimisation purpose. Different architectures of CNN and CCNN are implemented for contrast function estimation. Both simulation results and application to the STAR*D (Sequenced Treatment Alternatives to Relieve Depression) trial indicate that the proposed methods outperform penalised least square estimator.