Prediction of response to cardiac resynchronization therapy using a multi-feature learning method