| 摘要: | 近年來，隨著深度學習的技術迅速成長及其在許多人工智慧的議題上的取得明顯進步，使得問答系統獲得許多的關注與研究，例如史丹佛大學建構的 SQuAD 資料集吸引了許多國際自然語言處理研究團隊（例如Google與Microsoft團隊）的興趣。這一年來，如何將AI運用於生醫領域，成了最熱門的話題之一。 2019 年Google公開了自己建構的 PubMed QA資料集，並建構一個線上平台評估使用者在這個資料集上的效果。儘管問答系統在生醫自然語言處理上開始得到許多關注，但無論是Google 的 PubMed QA 資料集或是由美國衛生研究院贊助的國際生醫問答系統競賽BioASQ，主要都是以生醫論文為主的資料集。截至目前為止，仍非常缺乏以臨床病歷作為資料集的問答系統。臨床病歷一直以來都是生醫自然語言處理當中非常重要的研究領域，例如美國衛生研究院與美國梅奧醫院一直以來致力於參與及推廣臨床病歷上的各項自然語言處理任務，例如：去識別化、危險因子辨識（risk factor detection）與家族病史偵測（family history extraction）等。到近年來 IBM 更是提出一個 GAMENet （Graph Augmented MEmory Networks）模型來提供醫生在對病人用藥上的建議。然而這些研究依然多是以分類問題為主，缺乏臨床病歷的問答系統，例如：這位病人有糖尿病症狀，除了可能會有心血管疾病外，可能還會有哪些併發症？有哪些藥物可以抑制降低糖尿病的症狀，並且比較不容易引起心臟病的併發症？相較於過去在臨床病例上的自然語言處理任務，臨床文獻的問答系統對於醫生做診斷時可以有更即時地幫助，也更加具有挑戰性。在本計畫當中，本團隊將在共同主持人黃彥霖醫師的協助下，建構一個臨床病例的問答語料庫與系統。預期建構的問答語料庫會以病人的症狀、疾病、併發症與用藥建議的問題為主，並用來協助醫師寫住院病摘，本計畫預期為執行三年之計畫，第一年預期利用自然語言問句生成模型建構一臨床病例問答的資料集。第二年則延續前一年之資料利用深度學習技術建構一個疾病與藥物的知識圖以及來輔助深度學習預測症狀、疾病、併發症與用藥上的關聯性。最後一年，除黃醫師外，預期邀請其他國內醫師合作，建構臨床診斷輔助問答系統，並將積極推廣本計畫所得出之資料集與標準，促進國內醫療的進步。 ;In recent years, deep learning technologies growth rapidly and have widely applied in many artificial intelligence topics. The researchers gradually increase attention to question answering systems. For example, Stanford University constructed the SQuAD (Stanford Question Answering Dataset) dataset and has attracted the interest of many international natural language processing (NLP) research teams, such as Google and Microsoft teams. In biomedical NLP, the QA system also receives much attention. For example, in 2019, the Google research team released their own PubMed QA (Question-Answering) dataset and constructed an online platform to evaluate users' performances on the dataset. However, both Google's PubMed QA dataset and the International Biomedical Question Answering System competition, BioASQ (Biomedical Semantic Indexing and Question Answering) dataset, which is sponsored by the National Institute of Health, NIH) are from a collection of biomedical literature. To our acknowledgment, there is no available clinical QA dataset and system. The clinical NLP tasks are always a critical research area in NLP. For example, the NIH and American Mayo Clinic hospitals have been committed to participating in and promoting various clinical NLP tasks, like de-identification, risk factor detection, and family history extraction. In recent years, IBM has proposed a GAMENet (Graph Augmented Memory Networks) model to assist diagnosis by recommending the patient's medications for doctors. However, most of these studies are usually classification problems but do not include a clinical QA system. For example, the patient has diabetes symptoms. In addition to cardiovascular disease, what other complications may the patient have? What drugs can suppress the symptoms of diabetes and are less likely to cause complications of heart disease? Compared with the previous NLP tasks, the QA system in clinical literature is more challenging but can help doctors more immediately in diagnosis. In this proposal, we plan to develop a clinical QA dataset and system. Our QA corpus focuses on the patient's symptoms, diseases, complications, and medication recommendations, and will be used to assist doctors in writing the clinical record. This plan is expected to be a three-year plan. In the first year, we plan to use the QA generation model to construct the dataset. In the second year, the dataset from the previous year is used to develop deep learning technology to construct a Knowledge Graph of diseases and drugs, which will be used to predict the relations between symptoms, diseases, complications, and medication. In the last year, we plan to construct a QA system and collaborate with domestic doctors. Besides, the results of this project are expected to become the international standard benchmark for clinical QA. |
