Towards deep anchor learning

Mads A. Hansen, Karl Øyvind Mikalsen, Michael Kampffmeyer, Cristina Soguero-Ruíz, Robert Jenssen

Abstract: Anchor learning is a promising framework for electronic health record phenotyping and for predicting clinical state variables. In particular, anchor learning alleviates the problem of having to manually annotate (label) a large dataset (time consuming and expensive) by transforming certain key clinical observations (anchors) into noisy labels. Thus far, the predictive models in anchor learning have utilized rather simple features for the analysis. We hypothesize that anchor learning may benefit hugely from more powerful feature extraction. Recent advances in deep learning have illustrated neural networks' exceptional ability to learn powerful feature representations from data. In this paper, we propose the deep anchor learning framework, which uses a deep autoencoder to obtain stronger features for use in anchor learning. A support vector machine trained on the autoencoder features achieves state-of-the art performance on the task of postoperative delirium prediction from anchors.