Probing clinical NLP models for structural bias