藥物推薦在醫療資訊的應用領域是一項重要任務。先前的方法論都沒有善加利用 診療紀錄之間的醫療代碼的相似性來促進學習，並且過度強調單一病患的歷史診療紀 錄，而沒有妥善利用大量只有一次診療紀錄的病患資料。同時，近年的方法論中大多 數需要依靠外部知識的協助或是複雜的模型設計來提促進表現，使模型的適用範圍愈 趨狹隘，並且大多數研究都只以 MIMIC-III 資料集進行驗證。本研究提出一個能夠有效 利用所有看診紀錄的方法論 GSVEMed，並且使用兩個電子醫療紀錄資料集執行實驗， 強調以資料集本身的學習促進表現而不依賴外部知識，在結構簡單的情況下於 MIMIC- III 資料集取得與最先進作法相抗衡且在私人資料集明顯超過最先進作法的表現，並且 根據不同加護病房類型與醫院內科科別進行分析。;Medication recommendation is an important task in healthcare informatics. Previous methodologies have not effectively utilized the similarity of medical codes between visit records to facilitate learning. They have also overly emphasized the historical visit records of individual patients, without properly utilizing a large amount of patient data that consists of only one visit record. Additionally, most recent methodologies have relied on external knowledge or complex model architecture to improve performance, making the scope of application increasingly narrow. Furthermore, most studies have only validated their approaches using MIMIC-III dataset. This study proposes a method called GSVEMed that effectively utilizes all visit records. We conduct experiments using two electronic medical record (EMR) datasets, emphasizing performance improvement through learning from the datasets themselves rather than relying on external knowledge. Under the condition of a simple architecture, GSVEMed achieves performance comparable to state-of-the-art approaches on MIMIC-III dataset and significantly outperforms them on our private dataset. This study also conducts analyses based on different types of intensive care units of MIMIC-III and internal medicine departments of the private dataset.
