請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/70297
標題: | 緊急醫療派遣對話中報案人情緒的自動辨識 Automatic Caller Emotion Recognition During 911 Calls for the Emergency Medical Dispatch |
作者: | Tzu-Chun Hsieh 謝子鈞 |
指導教授: | 陳柏華(Albert Y. Chen) |
關鍵字: | 情緒辨識,緊急醫療系統,派遣溝通,支持向量機, Emotion Recognition,Emergency Medical Services,Dispatch Communication,Support Vector Machines, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 如何快速辨識到院前心肺功能喪失 (out-of-hospital cardiac arrests, OHCA)的病患是很重要的醫療與公眾健康問題,因為病患到院前心肺功能喪失的存活率,將隨時間推移而迅速下降,而緊急報案電話 (119) 為接觸緊急事件報案的前線,由派遣員聽取報案人提供之資訊並派遣救護車,如果在報案前線便能精準且快速辨識是否為緊急的患者,將大大提升病患的存活率,但報案人面對緊急事件時往往伴隨著緊張、害怕等情緒,這些情緒會影響後續派遣員判斷病患嚴重程度的能力,進而導致延誤就醫或處置失當,因此了解及辨識報案人情緒是相當重要的議題,也是本研究的核心。
若要全盤了解報案人情緒,則需要聆聽相當多報案錄音並進行分析,但以全人工聆聽報案音訊,將會是一件費時耗力的事,且以人耳聆聽難免參雜些微主觀判斷意識,且判斷標準會隨著聆聽時間長度越長而失準。因此,本研究使用機器學習 (Machine Learning, ML),快速建立精確與穩定的情緒辨識模型,使報案者的情緒分析能夠更加迅速、穩定和客觀,也使得眾多報案錄音得以進行分析、研究。本研究與台大醫院急診部合作,並運用2011年後的428個台北市消防局的真實報案資料進行研究與分析,期望未來能提供民眾更好的緊急醫療派遣品質,也幫助消防局進行資料歸檔與整理,使報案錄音更能發揮其應有之價值,也提供派遣員未來在訓練報案應對流程時加入未發現之改善建議。 先前研究中,雖有提出自動情緒辨識的方法,卻鮮少運用在緊急派遣電話中。緊急派遣電話錄音有說話急促、大小聲不一、派遣員與報案人對話交雜與雜訊多等特徵,比一般所用之音訊更增添困難。本研究建立一套完整的處理與辨識過程,有效克服眾多困難,最後結果顯示我們所提出的方法,能穩定快速且精確的辨識情緒的穩定程度。分析過程中,我們先將音訊噪音去除、自動將派遣員音訊分離再進行音訊特徵截取,其中抽離出梅爾頻譜係數 (Mel-Frequency Cepstral Coefficients, MFCCs)並保留其中12維度的特徵向量,最後送入支持向量機 (Support Vector Machines , SVM) 進行高維度特徵的情緒辨識。 其中在情緒真值的部分,我們使用了情緒與內容配合分數 (Emotional Content and Cooperative Scores, ECCS),以人工聽取錄音檔進行ECCS階層分類,情緒與內容配合分數是以報案人情緒穩定程度與對派遣員之配合程度作為衡量依據,其中階層四、五表示報案人無法配合派遣員進行正常的緊急醫療派遣對話。在我們的第一個資料集中,只有25位 (7%)報案人被歸類為階層四或五;而在第二個資料集中,則有11位 (18%)報案人被歸類為階層四或五,顯示緊急醫療派遣溝通中報案人情緒趨於穩定。 此研究結合機器學習與緊急醫療派遣,並使用真實資料進行測試與分析,結果顯示使用這一套流程,我們能以93.15%的準確率成功分類報案人情緒的穩定程度,並協助派遣員進行緊急醫療派遣,且其中說話者分離 (Speaker separation)的方法將只占總體派遣錄音約39%的報案人音訊分割出來,減少派遣員情緒對本研究報案人情緒辨識的影響,並提高整體情緒辨識準確率達0.56%。未來期望能提供社會大眾更好的緊急醫療派遣系統,提升派遣速度且降低傷亡率。 The emergency medical dispatch (EMD) system provides assistance to the people under emergency medical situations. In the EMD system, the caller needs to provide the patient’s condition and the incident occurred to help the dispatcher know the situations. The dispatcher needs to provide appropriate assistance according to the situation such as to dispatch the appropriate type of ambulance, and teach the caller to perform Cardiopulmonary Resuscitation (CPR). However, the EMD communication exists some barriers between the caller and the dispatcher such as emotion, language and knowledge. This research focus on the study of the emotion barrier to enhance a more efficient and effective dispatch, especially the emotion status of callers. The caller could get emotional when facing non-routine incidents. Previous works study the emotion status of callers by listening to the EMD recordings one by one. However, the whole process was tedious and time-consuming. In this study, a novel method for the emotion classification using a series of methods is proposed. The methods include a preprocessing technique to help clean the raw recordings, speaker separation to extract the caller audio from the original audio, standardization to prepare data for classification, and the Mel-Frequency Cepstral Coefficients (MFCCs) features and the Support Vector Machines (SVM) to classify the emotion status of callers. A case study in Taipei, Taiwan was conducted with two datasets. In the study, there are 428 actual audio recordings/signals of emergency calls collected after 2011, provided by the Taipei City Fire Department. In order to validate the performance of the proposed model, the emotion status of caller in the EMD communication is labeled manually through an emotion standard, Emotional Content and Cooperative Scores (ECCS). This standard is used to classify the emotion status of callers in the EMD communication. In dataset 1, there are 25 (7%) recordings marked emotional unstable which represent the ECCS level larger than 3. In dataset 2, there are 11 (18%) recordings marked emotional unstable. This shows a relatively stable emotion of caller in the EMD communication. Results suggest that the proposed method has potential to recognize emotion with an accuracy of 92.59% and can be further improve to 93.15% by implement speaker separation, which helped us better understand the barrier and help increase the overall performance in the communication of pre-hospital EMD system. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/70297 |
DOI: | 10.6342/NTU201803319 |
全文授權: | 有償授權 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf 目前未授權公開取用 | 1.75 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。