Towards Understanding ASR Error Correction for Medical Conversations