We introduce Doctor XAvIer \textemdash a BERT-based diagnostic system that extracts relevant clinical data from transcribed patient-doctor dialogues and explains predictions using feature attribution methods. We present a novel performance plot and evaluation metric for feature attribution methods \textemdash Feature Attribution Dropping (FAD) curve and its Normalized Area Under the Curve (N-AUC). FAD curve analysis shows that Integrated Gradients outperforms Shapley values in explaining diagnosis classification. Doctor XAvIer outperforms the previous diagnostic system with 0.97 F1-score in named entity recognition and symptom pertinence classification and 0.91 F1-score in diagnosis classification.