現今,人工智慧在醫療領域的應用隨著醫療需求的增加也越加受到重視。在各種醫療行為中,門診對話是大多數患者在尋求醫療幫助時都會經歷的過程。由於患者隱私問題,門診對話和患者醫療記錄的收集受到許多限制。此外,研究門診對話的研究人員通常無法公開他們的數據集。因此,以往的工作大多以在線醫療社區的諮詢對話作為研究資料,但這些諮詢對話與門診對話仍有不小的差別。我們與台大醫院(NTUH)家庭醫學部合作,為研究獲取了一些關於門診對話和患者醫療記錄的數據。我們使用基於Transformer的模型進行自動化的門診對話摘要。在訓練的過程中,我們引入了很多外部的醫學數據集來幫助模型學習醫學的術語和知識。由於我們提出的方法是透過分段對話再進行摘要,因此模型可以處理比較長的門診對話。此外,我們還使用NTUH的資料集來訓練文本風格轉換模型來模仿醫師做的醫療筆記。實驗結果顯示我們方法生成的門診對話摘要具有一定的參考價值。
Nowadays, the application of artificial intelligence in the medical field is getting more and more attention with the increase in medical demand. Among various medical practices, outpatient conversation is a process that most patients experience when seeking medical assistance. Due to patient privacy concerns, the collection of outpatient conversations and patient medical records is subject to many limitations. Furthermore, researchers studying outpatient conversations are often unable to make their datasets public. Therefore, most of the previous work used consultation conversations in online medical communities as research materials, but these consultation conversations are still quite different from outpatient conversations. We collaborated with the Department of Family Medicine of National Taiwan University Hospital (NTUH) to obtain some data on outpatient conversations and patient medical records for the study. We use Transformer-based models for automatic summarization of outpatient conversations. During the training process, we introduced many external medical datasets to help the model learn medical terms and knowledge. Since our proposed method performs summarization through segmented conversations, the model can handle relatively long outpatient conversations. Additionally, we use the NTUH dataset to train a writing style conversion model to mimic medical notes made by physicians. The experimental results show that the outpatient dialogue summaries generated by our method have a certain reference value.