Dropout Feature Ranking for Deep Learning Models

Deep neural networks are a promising technology achieving state-of-the-art results in biological and healthcare domains. Unfortunately, DNNs are notorious for their non-interpretability. Clinicians are averse to black boxes and thus interpretability is paramount to broadly adopting this technology. We aim to close this gap by proposing a new general feature ranking method for deep learning. We show that our method outperforms LASSO, Elastic Net, Deep Feature Selection and various heuristics on a simulated dataset. We also compare our method in a multivariate clinical time-series dataset and demonstrate our ranking rivals or outperforms other methods in Recurrent Neural Network setting. Finally, we apply our feature ranking to the Variational Autoencoder recently proposed to predict drug response in cell lines and show that it identifies meaningful genes corresponding to the drug response.