Weakly Supervised Medication Regimen Extraction from Medical Conversations

Automated Medication Regimen (MR) extraction from medical conversations can\nnot only improve recall and help patients follow through with their care plan,\nbut also reduce the documentation burden for doctors. In this paper, we focus\non extracting spans for frequency, route and change, corresponding to\nmedications discussed in the conversation. We first describe a unique dataset\nof annotated doctor-patient conversations and then present a weakly supervised\nmodel architecture that can perform span extraction using noisy classification\ndata. The model utilizes an attention bottleneck inside a classification model\nto perform the extraction. We experiment with several variants of attention\nscoring and projection functions and propose a novel transformer-based\nattention scoring function (TAScore). The proposed combination of TAScore and\nFusedmax projection achieves a 10 point increase in Longest Common Substring F1\ncompared to the baseline of additive scoring plus softmax projection.\n