Attention-based multimodal fusion with contrast for robust clinical prediction in the face of missing modalities.