透過機器學習之文本分析在緊急醫療派遣中提升識別重大創傷報案的準確率

Abstract

準確識別重大創傷緊急報案，可以派遣與案件嚴重程度相應之救護車，提供病患更好的救護，以降低致命風險。根據本研究之了解，目前極少研究應用機器學習方法於緊急報案派遣階段實行創傷案件分類；因此，本研究旨在構建文本分類模型，以實現到院前啟動之重大創傷報案電話的自動化識別，並且找出代表創傷案件的顯著特徵。本研究的分析資料為隨機抽樣2018年台北市車禍報案電話，共114個車禍報案電話，其中72筆資料為非重大外傷，42筆為重大外傷案件。報案內容的文字分詞是由中研院中文知識和信息處理小組開發的分詞器所執行。此外，本研究使用TF-IDF實現特徵選擇，並採用伯努利樸素貝葉斯作為分類模式。為了評估模型的分類結果，本研究邀請台北市與新北市共6位緊急救護派遣員依據自身經驗對114筆案件進行判斷，並在判斷時給予答題的確定程度。本研究透過模型自動選出的特徵結合專業派遣員提供的關鍵詞，提出了半自動特徵選擇模型。此創新模型搭配上客製化的規則判斷設計，能夠在派遣員應答信心程度較低時，給予相對較高準確度的分類答案。這是第一個驗證在到院前急救醫療服務期間，透過機器學習模型識別到院前啟動之重大創傷與非重大創傷報案電話的研究。本研究設計的模型不僅提出識別創傷之重要關鍵字，並且能夠在派遣員信心較不足時，提高創傷案件分類的準確率；因此，此研究領域在創傷醫療派遣中，有許多探討之空間。

Accurate identification of major-trauma emergency calls enables the dispatch of an appropriate ambulance type to improve pre-hospital care, thereby potentially reducing risk of fatality. To the best of the author’s knowledge, very few studies have focused on classification of trauma cases based on machine learning frameworks. This study aims to construct a text classification model to achieve identification of Pre-hospital Activated Major Trauma (PAMT) and non-PAMT calls, and to discover significant features of these calls. In the analysis, 114 car accident calls in Taipei, Taiwan are randomly sampled from 2018. Among the calls, 72 are cases without pre-hospital major trauma activation and 42 are PAMT cases. The complete semantic word segmentation of the calls was performed via the Chinese Knowledge and Information Processing Tagger (CKIP Tagger). Moreover, the Term Frequency-Inverse Document Frequency (TF-IDF) was implemented to select features and the Bernoulli Nave Bayes was employed as the classification model. To evaluate the results, 6 Emergency Medical Dispatchers (EMDs) in Taipei City and New Taipei City participated in an interview to provide their judgment of PAMT and non-PAMT to the calls and whether they were certain when they answered. In this study, with the features automatically selected by the model and the keywords provided by the dispatchers, a semi-automatic feature selection model is proposed. This innovative model and combined with a customized rule-based judgement holds a better classification performance when dispatchers are uncertain about the particular case. This is the first study to validate the recognition of trauma calls by machine learning frameworks. The important keywords were detected, and the classification outcome was enhanced by the text classification model. Thus, this study area is worth of further investigation and development in trauma medical dispatch.