Logistic regression with a large number of features compared to available labels presents numerous challenges for learning. We present screening rules that safely remove features from the sparse logistic regression with L0-L2 regularization before solving the problem. The screening rules are based on the Fenchel dual of strong conic relaxations of the sparse logistic regression problem. Numerical experiments with real and synthetic data suggest that a high percentage of the features can be effectively and safely removed apriori, leading to substantial speed-up in the computations.