人類大腦功能對應知識庫:研究方法、系統與相關應用

Abstract

近幾十年來，科學家已在各個層次上進行神經活動和腦功能為目標的神經科學研究，並取得了歷史上任何時代都無可比擬的重大發展。而近20年來科學家廣泛的仰賴以非侵入式的功能性磁振造影(fMRI) 進行大腦功能研究，因此以fMRI研究大腦功能的文獻大量的成長，成為一個極具有重要價值的參考內容。然而，閱讀大量的文獻是一件相當費時費力的工作，往往都是需要再經過自己的吸收消化才能成為具有價值的知識，因此，若有一套資訊自動擷取系統可以由文獻中取得有用的參考資訊，譬如由文獻中得知不同實驗刺激對大腦區域的反應以及大腦區域與功能的關係等資訊，除可大幅降低閱讀所需的人力與提高閱讀時的效率，相信可以對瞭解大腦功能與實驗設計提供極大的助益。 本篇論文之目的主要是發展以及應用資訊擷取技術由fMRI文獻中進行資訊擷取與探勘，進而建構人類大腦功能對應知識庫。在第一部份的資訊擷取研究中，首先利用統一醫學語言系統(Unified Medical Language System)建立一個廣義的階層化概念式大腦功能字典 (generalized hierarchical concept-based dictionary of brain functions)，再搭配以字典以及規則為基礎的混和式資訊擷取演算法來標記大腦功能相關字彙並將之對應至相關的概念(Concept)。在本研究中所建立的大腦功能字典為目前第一個廣義的階層化概念式大腦功能字典，而本研究中所發展的資訊擷取系統更是目前第一個可以從大腦功能文獻中自動擷取大腦功能與實驗任務辭彙的系統，其精確度(Precision)和召回率(Recall)皆可達到相近於專家的正確性。 第二部分的研究，發展了一套大腦功能對應知識庫(BrainKnowledge: a human brain function mapping knowledge-base) 。利用第一階段所研發的資訊擷取技術進行大量文獻分析，並將擷取的資訊建構出大腦結構與功能的共現模型(co-occurrence model)，本研究更進一步整合fMRI 實驗與文獻分析，以文獻探勘結果輔助學者解釋實驗的結果，提供研究大腦認知功能的學者以更快更有效率的方式學習，以便能更進一步對人類大腦功能研究做更詳細的分析。 總結而言，本論文成功地發展一套大腦文獻資訊擷取系統，由文獻中標記出大腦區域、功能與實驗任務辭彙，並將此資訊用於輔助大腦實驗的解釋，相信對於大腦功能研究會有相當的助益。

With a rapid progress in the field, a great many functional magnetic resonance imaging (fMRI) studies are published every year, to the extent that it is now becoming difficult for researchers to keep up with the literature, since reading papers is extremely time-consuming and labor-intensive. Thus, automatic information extraction has become an important issue. The purpose of this dissertation is to develop and validate an information extraction algorithm for extracting information from the fMRI literature. It is divided into two parts. First, we developed a generalized hierarchical concept-based dictionary of brain functions for named entity extraction based on the Unified Medical Language System (UMLS), which integrates many terminologies such as MeSH, Psychological Index Terms and similar vocabulary sources. Second, a hybrid method that combined a dictionary and a rule-based approach for recognizing and classifying concepts related to human brain studies. To the best of our knowledge, this is the first study to extract brain functions and experimental tasks from the fMRI literature automatically. The generalized hierarchical concept-based dictionary of brain functions we have developed is the first generalized dictionary of this kind. It can be helpful for further studies in text mining, as can algorithms for automatic retrieval of brain functions and their hierarchical relationships for cross-referencing. The precision and recall of our information extraction algorithm was on par with that of human experts. Our approach presents an alternative to the more laborious, manual entry approach to information extraction. In addition, to demonstrate the possible applications of the extracted terms, we present a human brain function mapping knowledge-base system (BrainKnowledge) that combines fMRI datasets with the published literature in a comprehensive framework for studying human brain activities. BrainKnowledge not only contains indexed literature, but also provides the ability to compare experimental data with those derived from the literature. In this dissertation, we will describe BrainKnowledge, which provides concept-based queries organized by brain structures and functions and also mines results to support or explain the experimental fMRI results, and we will present application examples with the studies of affect, and studies of language to illustrate its capabilities. In summary, this dissertation successfully demonstrated the ability of information extraction from the fMRI literature. Furthermore, BrainKnowledge, which combines fMRI experimental results with Medline abstracts, may be of great assistance to scientists not only by freeing up resources and valuable time, but also by providing a powerful tool that collects and organizes over ten-thousand abstracts into readily usable and relevant sources of information for researchers.