請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91996
標題: | 基於混合深度模型之具可解釋性的心跳驟停與心肺復甦預警系統 An Explainable Deep Early Warning System based on Hybrid Model for Cardiac Arrest and Cardiopulmonary Resuscitation |
作者: | 鄧遠祥 Yuan-Xiang Deng |
指導教授: | 傅立成 Li-Chen Fu |
關鍵字: | 急診室,心跳驟停,預警系統,多模態模型,時序資料, Emergency Department,Cardiac Arrest,Early Warning System,Multi-modal model,Time Series Data, |
出版年 : | 2024 |
學位: | 碩士 |
摘要: | 急診室壅塞是全球公共衛生面臨的重要挑戰之一,而院內心跳停止為其最嚴重的併發症之一。為了提高醫療資源的有效分配,以及增加患者的生存率,眾多研究致力於早期檢測心跳停止。本研究提出了一種基於深度混合模型的心跳停止早期預警系統。我們使用了來自台大醫院急診室的數據,其中包含心跳、血氧、血壓等生理資訊的時間序列數據。由於急診室收集的時間序列資料中,不規律採樣間隔和大量缺失值的普遍存在,因此我們需找出合適的補值方式。本研究評估了先前用於處理不規則時間序列的多任務高斯過程是否適用於我們的資料。此外,本研究還整合了病患的基本資訊,以利用多模態資料提供更豐富的信息表示。
我們的系統測試測試在10839個樣本、不平衡比例為1%的測試集上,AUPRC和AUROC分別達到0.5178和0.9388。此外,本研究已嘗試了的心肺復甦檢測這項更具挑戰性的任務。心肺復甦是心跳停止的一個子集,相應的數據量較有限,另外不預期的心跳停止患者的生命徵象測量數據沒有瀕臨死亡患者的完整,因此在排除簽屬不急救病患後,預測心肺復甦病患是更困難的。儘管檢測需要心肺復甦的患者在臨床上具有更高的價值,但是在心肺復甦檢測任務中,使用深度學習的相關文獻相對稀缺。因此,本研究開展了對心肺復甦檢測任務的初步探索。在心肺復甦任務中,實驗結果顯示我們的系統在包含10762個樣本、不平衡比例為0.3%的測試集中,AUPRC和AUROC分別達到0.0604和0.7438。 The congestion of emergency rooms in a hospital is among the most critical public health issues globally, with in-hospital cardiac arrest being one of the most severe complications. An increasing amount of research is dedicated to early detection of cardiac arrest, allowing for more efficient allocation of medical resources and significantly improving patient survival rates. This thesis proposes an early warning system for cardiac arrest based on a deep mixture model. This study leverages the data collected from the emergency department of National Taiwan University Hospital, which is a kind of time series data comprising physiological information such as heartbeat, blood oxygen levels, and blood pressure. Irregular sampling intervals and significant missing values are common in the time series data collected from the emergency rooms, necessitating the identification of suitable imputation methods. This study assessed whether the multitask Gaussian process, previously employed in handling irregular time series in past research, is applicable to our dataset. It is also worth noting that this research incorporates basic patient information, employing multimodal data to learn representations enriched with abundant information. Our system exhibited promising performance in the cardiac arrest detection task, as evidenced by experimental results on a test set comprising 10,839 samples with an imbalance ratio of 1%. The AUPRC and AUROC achieved were 0.5178 and 0.9388, respectively. Additionally, this study ventured into the more challenging task of detecting cardiopulmonary resuscitation (CPR). CPR is a subset of cardiac arrest with a limited dataset, and vital sign measurements for patients experiencing unexpected cardiac arrest are less comprehensive compared to those nearing death. Therefore, predicting CPR patients, excluding patients who sign Do-Not-Resuscitate, is more challenging. Despite the higher clinical value associated with detecting patients requiring CPR, there is a scarcity of literature utilizing deep learning in CPR detection tasks. Therefore, this study embarks on an initial exploration into the task of CPR detection. In the CPR detection task, experimental results revealed that our system achieved an AUPRC of 0.0604 and an AUROC of 0.7438 in a test set containing 10,762 samples with an imbalance ratio of 0.3%. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91996 |
DOI: | 10.6342/NTU202400299 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf 目前未授權公開取用 | 3.12 MB | Adobe PDF | 檢視/開啟 |
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。