Towards More Generalizable Machine Learning: Improving Model Robustness Against Clinical Event Sequence Shifts