Causal Reinforcement Learning based Agent-Patient Interaction with Clinical Domain Knowledge