皮膚病理非感染性全身性水泡病診斷決策支援系統
A Dermatopathological Diagnostic Decision Support System for Non- Infectious Generalized Blistering Diseases
中文摘要
忙碌的臨床醫師在面臨病患時,常常要及時下正確的診斷以便盡快給予病人治 療,本研究藉由建立一個皮膚病理診斷決策協助系統,協助醫師在皮膚科的非 感染性全身性水泡病(Generalized Blistering Diseases - GBD)領域做臨床診斷決策。
並透過Internet 來使用 WWW 使用者介面,而不受作業系統及地理區隔之限制。
由於臨床皮膚病理學推論機制上的不確定性,在GBD 領域,運用 Multi-
membership Bayesian Formulation 的機率性推論知識表現法,來發展將醫學知識 轉化為知識庫,而能為電腦所用的知識工程過程,以建造該領域醫學知識庫。在 這個知識工程的過程中,主要的人力包括皮膚病理學專家、協調主持人、知識工 程師及程式設計人員。在整個計畫的前幾個月首先發展出WWW 介面架構的介 殼系統(System Shell),此介殼系統包括推論引擎與使用者介面等,接著以近六 個月的時間隔週開會,建造出資料辭典與知識庫,在系統整合完成後並進行系 列的測試評估。
在建造醫學知識庫方面,GBD 領域內的每個疾病,以一個 Bayesian 框架(Frame) 來代表,每個框架由該疾病的診斷條件(Findings)組成,並設定每個框架的 Apriori 與每個診斷條件的 True Positive Rate (TPR), False Positive Rate (FPR),
由知識工程人員聚會討論,將皮膚病理學專家的知識數據化、規則化,來建立這 些診斷條件, Apriori, TPR, FPR,並從學術文獻中搜尋參考相關之數據及知識,
以協助知識工程之完整性及客觀性。在知識工程完成後再請不同的皮膚病理學專 家協助,讓知識庫中可能有爭議性的項目,透過Delphi method 達成共同之客觀 結論。在醫院中或國際學術文獻期刊上發表的臨床病理討論病例,被選用為臨床 測試對像,並針對知識庫做最後的修正。
在Bayesian Formulation 的機率性推論知識表現法方面,是假設在每個框架內的 診斷條件之間是條件獨立(condition independent)的,以避免系統會對相關診斷條 件產生機率推估過高(overconfidence)現像,本研究則對非條件獨立(non-
independence)的診斷條件,採取調整 TPR 與 FPR 的做法,以提高最後機率計算 的正確性。
這個系統知識庫總共建造了11 個代表 GBD 領域疾病的框架,資料辭典中共有 90 個診斷條件或疾病項目,每個框架約有 5 到 10 個代表臨床、光學顯微鏡及免 疫病理學方面的診斷條件,共171 個代表 Apriori, TPR, FPR 的機率數據。
從學術文獻期刊中取得28 個臨床病理病例做為測試對像,其中 25 個測試病例 能正確的將最可能的疾病機率計算出來,即使不是本系統的11 個疾病項目,而 是其中疾病的變化型,本系統所計算出的診斷,亦為最接近其原型的可能疾病,
參與計畫的皮膚病理學專家認為這個系統符合可用性。
建造web-based 介殼系統的診斷決策支援系統有幾項優點,(1)在不同地域的眾 多使用者,可以同時維護同一醫學知識庫,橫跨不同國家的醫學專家可以共同 合作建造同一系統。(2)只要具備能連接上 Internet ,並擁有 WWW 瀏覽器的電 腦環境, 就可以使用本系統,並不需要額外購買或裝置軟硬體。(3)不同醫學科 別領域的多種醫學知識庫,能夠共同建置在同一系統中,做到跨醫學科別領域 的診斷決策支援與臨床教學用途。
建置機率性推論的web-based 皮膚病理學診斷支援系統,除了在診斷決策支援 與臨床教學用途有所幫助之外,在發展知識工程的過程中,對這個醫學領域的 知識能夠有更清楚的認識,對將來發展皮膚科的證據醫學亦將有所助益。
英文摘要
In recent decades of "information explosion", it is difficult for practicing physicians to keep up to date with all the information for making optimal clinical judgements.
The aim of this project tried to build a dermatopathological decision support system to help physicians to make optimal clinical decisions.
In lights of the uncertainties involved in the dermatopathological reasoning process, we proposed a probabilistic knowledge representation under the scope of
"non-infectious generalized blistering diseases". In order to create the knowledge base, a technique termed “knowledge engineering” was implemented. This knowledge engineering process included one dermatopathological expert, one coordinator and one knowledge engineer. This group met every two weeks during a six-month period to build up a data dictionary and knowledge base.
Programming engineer developed the inference engine and user interface in a web-based architecture. Vigorous evaluation process was performed after all parts of the system were put together. Both theoretical performance indices and clinical performance are to be tested. Published CPC (clinical-pathological conference) cases and patients charts are used for the final evaluation.
This prototype knowledge base represents 11 disease frames with 88 findings in the data dictionary. Each disease frame consists of 5~10 findings, which
represents clinical, light microscopic and immunopathologic clues for the diagnosis of a disease.
There are several advantages with the web-based system shell and the user interface. First, multiple users in different geographical areas can update their knowledge base in the same time, so the international cooperation is feasible.
Second, those who like to use the system need only a Internet access to do so. No additional software needs to be sent or installed. Third, multiple knowledge bases in different medical domains can be incorporated into the system for cross-
domain diagnostic decision support and clinical teaching.
This project proposes to apply probabilistic reasoning to a web-based
dermatopathological diagnostic decision support system. The system can be useful in both clinical teaching and diagnostic decision support. In the process of knowledge engineering, we also felt that the domain medical knowledge was better clarified and may contribute to the future of evidence-based medicine in dermatology.