從電子病歷中擷取問題列表

Abstract

問題列表方式的病歷記載近年來已漸漸成為醫療記錄方式的主流。然這種表列式的問題分析記錄方式，還是不能取代傳統的病史敘述。問題列表式的紀錄主要包含主觀的症狀、客觀的檢驗檢查結果、包含診斷的評估及處置計畫四大部分。多數的問題列表是從病史敘述中擷取產生的，重點在於症狀與診斷關聯組合的產生。本研究利用某教學醫院一年期的出院病歷摘要之資料庫，嚐試建立症狀與診斷之關聯組合。 隨著每個人書寫的差異性在電子病歷上要自動化建立出診斷與症狀之間的關聯性是相當困難的，而且對於診斷與症狀之間的關聯性是屬於一種多對多的情況，若要使用人工來一個個建立這之間的關聯性會耗費相當大量的人力與時間，因此我們希望作一套系統以減輕醫生與整理病歷人員的負擔。 在面對龐大且複雜的電子病歷，我們以一套有結構性的方式去整理這些資料，並且從這些資料當中將病人的診斷與症狀以直觀的方式將其關聯性建立起來，而嚐試不需要經過太多的人工來建立當中的關聯性。 當病人將其主訴告訴醫生時，系統就可以經由這些病人所提供的敘述來擷取症狀，並且從這些症狀中對應到醫生可能下的診斷上，以協助醫生在下診斷時，可以參考可能患有的疾病或是其他同種症狀類似的診斷為何。在我們系統對於診斷的描述是從疾病本體論(Disease Ontology)的概念去描述，因此可以調整將這些從症狀相對應的診斷以概觀性較大範圍的描述或是較精確較小的描述去觀察診斷與症狀之間的關聯性。

Problem Listing had become an essential component of current medical record. In spilt of its popularity, problem listing is still a cumbersome task. A computer aid system will be most welcome if it can automatically extract subjective problems from patients’ description and may to a limited list of assessment. In the development of such system, it is important to build a comprehensive corresponding set between these two variables of patient’s problem and possible diagnosis. We need a medical database while covering one year period of discharge notes of a teaching hospital, to develop this knowledge set. The problems extended from chief complain were map to the ontology of main diagnosis of each case. One twenty selected cases were needed to develop the knowledge set, after domain expert intervention. Those selected training data set had tested for internal and external validation, the result showed the precision and recall of internal validation around 0.7 under low ontological resolution, and around 0.6 under higher resolution. The result of external validation were 0.5409 in precision, 0.6651 in recall, F-measure of 0.5966 under low resolution and 0.3851 in precision, 0.5923 in recall, F-measure of 0.4667 under high resolution. And the result of a common unimproved nature language processing tool were 0.398 in precision, 0.775 in recall, 0.652 in F-measure and the result of expert reviewer were 0.912 in precision, 0.788 in recall, 0.845 in F-measure. The result indicated the corresponding set between patient’s problem and diagnosis in our training set is too complicated to be resolved. By a one twenty selected training set in a high resolution setting, more intensive domain expert’s inversion is necessary to have a better result.