Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling